Added 11 `flashback-*` subsets from the Flashback internet forum.

CHANGELOG.md

@@ -5,6 +5,12 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
+## [v0.0.10] - 2026-07-26
+
+### Added
+
+- Added 11 `flashback-*` subsets from the Flashback internet forum: `flashback-dator`, `flashback-droger`, `flashback-ekonomi`, `flashback-hem`, `flashback-kultur`, `flashback-livsstil`, `flashback-ovrigt`, `flashback-politik`, `flashback-samhalle`, `flashback-sport`, `flashback-vetenskap`.
+
 ## [v0.0.7] - 2026-06-21
 
 ### Added

README.md

@@ -89,6 +89,50 @@ configs:
   data_files:
   - split: train
     path: data/cellar/*.parquet
+- config_name: flashback-dator
+  data_files:
+  - split: train
+    path: data/flashback-dator/*.parquet
+- config_name: flashback-droger
+  data_files:
+  - split: train
+    path: data/flashback-droger/*.parquet
+- config_name: flashback-ekonomi
+  data_files:
+  - split: train
+    path: data/flashback-ekonomi/*.parquet
+- config_name: flashback-hem
+  data_files:
+  - split: train
+    path: data/flashback-hem/*.parquet
+- config_name: flashback-kultur
+  data_files:
+  - split: train
+    path: data/flashback-kultur/*.parquet
+- config_name: flashback-livsstil
+  data_files:
+  - split: train
+    path: data/flashback-livsstil/*.parquet
+- config_name: flashback-ovrigt
+  data_files:
+  - split: train
+    path: data/flashback-ovrigt/*.parquet
+- config_name: flashback-politik
+  data_files:
+  - split: train
+    path: data/flashback-politik/*.parquet
+- config_name: flashback-samhalle
+  data_files:
+  - split: train
+    path: data/flashback-samhalle/*.parquet
+- config_name: flashback-sport
+  data_files:
+  - split: train
+    path: data/flashback-sport/*.parquet
+- config_name: flashback-vetenskap
+  data_files:
+  - split: train
+    path: data/flashback-vetenskap/*.parquet
 language_bcp47:
 - swe
 ---
@@ -99,7 +143,7 @@ language_bcp47:
 <!-- START README TABLE -->
 |              |                                                                                                                                                                |
 | ------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| **Version** | 0.0.7 ([Changelog](/CHANGELOG.md)) |
+| **Version** | 0.0.10 ([Changelog](/CHANGELOG.md)) |
 | **Language** | Swedish (sv, swe)                                                                                                  |
 | **License**  | Openly Licensed, See the respective dataset                                                                                                                    |
 | **Models**   | Currently there is no models trained on this dataset                                                                                                           |

data/flashback-dator/create.py

@@ -0,0 +1,144 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-dator` corpus.
+
+The source corpus has sentences scrambled for copyright/privacy reasons.
+Each sentence is emitted as a separate document.
+
+To run this script, set GIT_LFS_SKIP_SMUDGE=1 so uv does not download large
+Git LFS files while resolving the local Dynaword environment:
+
+    GIT_LFS_SKIP_SMUDGE=1 uv run data/flashback-dator/create.py
+"""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-dator"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-dator.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-dator.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    """Yield one record per sentence. Sentences are scrambled in the source corpus."""
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-dator/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 18929823,
+  "number_of_tokens": 512795188,
+  "min_length_tokens": 2,
+  "max_length_tokens": 65536,
+  "number_of_characters": 1527673015,
+  "min_length_characters": 1,
+  "max_length_characters": 131069
+}

data/flashback-dator/flashback-dator.md

@@ -0,0 +1,107 @@
+---
+pretty_name: 'Flashback: Computers & IT'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Computers & IT
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Computers & IT subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Computers & IT subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 18.93M
+- **Number of tokens (Llama 3)**: 512.80M
+- **Average document length in tokens (min, max)**: 27.09 (2, 65.54K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-dator_00000001",
+  "text": "Äh, överspännings skydd är onödigt, bytt hela datorn utan.",
+  "source": "flashback-dator",
+  "added": "2026-07-19",
+  "created": "2007-07-13, 2007-07-13",
+  "token_count": 26
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file. Språkbanken Text distributes the corpus with DOI `10.23695/0j57-m676`.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Computers & IT (updated: 2025-04-01). [Data set]. Språkbanken Text. https://doi.org/10.23695/0j57-m676
+
+```bibtex
+@misc{flashback_dator,
+  doi = {10.23695/0j57-m676},
+  url = {https://spraakbanken.gu.se/en/resources/flashback-dator},
+  author = {Språkbanken Text},
+  language = {swe},
+  title = {Flashback: Computers & IT},
+  publisher = {Språkbanken Text},
+  year = {2025}
+}
+```
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-dator/flashback-dator.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b31612ab7d83f6ef1b9aa286ad80444f8119dc4db4c6dbce8c0444182783f036
+size 1230583927

data/flashback-droger/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-droger` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-droger"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-droger.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-droger.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-droger/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 14515372,
+  "number_of_tokens": 403090735,
+  "min_length_tokens": 2,
+  "max_length_tokens": 10549,
+  "number_of_characters": 1171098003,
+  "min_length_characters": 1,
+  "max_length_characters": 22423
+}

data/flashback-droger/flashback-droger.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Drugs'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Drugs
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Drugs subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Drugs subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 14.52M
+- **Number of tokens (Llama 3)**: 403.09M
+- **Average document length in tokens (min, max)**: 27.77 (2, 10.55K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-droger_00000001",
+  "text": "Han gör vad som helst för wievs på sin kanal.",
+  "source": "flashback-droger",
+  "added": "2026-07-19",
+  "created": "2022-11-12, 2022-11-12",
+  "token_count": 14
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Drugs*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Drugs (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-droger/flashback-droger.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10591c5bf79c823909a20b4c8d3da96c2640c243f388d1f5b6686ad84aebe1f7
+size 956004586

data/flashback-ekonomi/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-ekonomi` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-ekonomi"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-ekonomi.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-ekonomi.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-ekonomi/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 8795695,
+  "number_of_tokens": 248006936,
+  "min_length_tokens": 2,
+  "max_length_tokens": 4641,
+  "number_of_characters": 723117973,
+  "min_length_characters": 1,
+  "max_length_characters": 8070
+}

data/flashback-ekonomi/flashback-ekonomi.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Economy'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 1M<n<10M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Economy
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Economy subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Economy subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 8.80M
+- **Number of tokens (Llama 3)**: 248.01M
+- **Average document length in tokens (min, max)**: 28.2 (2, 4.64K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-ekonomi_00000001",
+  "text": "Skulle jag belåna min fastighet, skulle jag ta ett reallån och betala realränta på det.",
+  "source": "flashback-ekonomi",
+  "added": "2026-05-30",
+  "created": "2008-03-29, 2008-03-29",
+  "token_count": 29
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Economy*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Economy (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-ekonomi/flashback-ekonomi.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec0f7e2e64bf88af9c1b7933bc40b6c853f459ba756d461153253a6f113fe543
+size 584779090

data/flashback-hem/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-hem` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-hem"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-hem.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-hem.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-hem/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 21921903,
+  "number_of_tokens": 617343364,
+  "min_length_tokens": 2,
+  "max_length_tokens": 6850,
+  "number_of_characters": 1822902453,
+  "min_length_characters": 1,
+  "max_length_characters": 9923
+}

data/flashback-hem/flashback-hem.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Home & Garden'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Home & Garden
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Home & Garden subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Home & Garden subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 21.92M
+- **Number of tokens (Llama 3)**: 617.34M
+- **Average document length in tokens (min, max)**: 28.16 (2, 6.85K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-hem_00000001",
+  "text": "Politikerna får ta och släppa på \" orörda \" fasader samt tillåta modernare tystare anläggningar än d[...]",
+  "source": "flashback-hem",
+  "added": "2026-07-19",
+  "created": "2018-07-26, 2018-07-26",
+  "token_count": 43
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Home & Garden*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Home & Garden (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-hem/flashback-hem.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878574cfc5377d51cc44729afb8b9d7102a304656e60bcff392bd44df5f7a348
+size 1451912226

data/flashback-kultur/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-kultur` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-kultur"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-kultur.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-kultur.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-kultur/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 30290455,
+  "number_of_tokens": 800007948,
+  "min_length_tokens": 2,
+  "max_length_tokens": 12196,
+  "number_of_characters": 2354106497,
+  "min_length_characters": 1,
+  "max_length_characters": 65535
+}

data/flashback-kultur/flashback-kultur.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Culture & Entertainment'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Culture & Entertainment
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Culture & Entertainment subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Culture & Entertainment subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 30.29M
+- **Number of tokens (Llama 3)**: 800.01M
+- **Average document length in tokens (min, max)**: 26.41 (2, 12.20K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-kultur_00000001",
+  "text": "Finns ju en hel del företag som trycker små upplagor och säljer åt dig.",
+  "source": "flashback-kultur",
+  "added": "2026-05-30",
+  "created": "2010-06-09, 2010-06-09",
+  "token_count": 26
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Culture & Entertainment*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Culture & Entertainment (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-kultur/flashback-kultur.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9a125aa90a76b99a4605e107f2f150ee94e23fd4067eac8a84d5c38caa6b48e
+size 1951741502

data/flashback-livsstil/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-livsstil` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-livsstil"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-livsstil.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-livsstil.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-livsstil/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 7218279,
+  "number_of_tokens": 189249970,
+  "min_length_tokens": 2,
+  "max_length_tokens": 7263,
+  "number_of_characters": 538261456,
+  "min_length_characters": 1,
+  "max_length_characters": 12003
+}

data/flashback-livsstil/flashback-livsstil.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Lifestyle'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 1M<n<10M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Lifestyle
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Lifestyle subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Lifestyle subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 7.22M
+- **Number of tokens (Llama 3)**: 189.25M
+- **Average document length in tokens (min, max)**: 26.22 (2, 7.26K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-livsstil_00000001",
+  "text": "The Horror!",
+  "source": "flashback-livsstil",
+  "added": "2026-05-30",
+  "created": "2013-04-12, 2013-04-12",
+  "token_count": 4
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Lifestyle*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Lifestyle (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-livsstil/flashback-livsstil.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4267b30a63172d5edeea130a5a27d8a762d15b1b9c5c3c9fef308e613741a577
+size 446804334

data/flashback-ovrigt/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-ovrigt` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-ovrigt"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-ovrigt.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-ovrigt.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-ovrigt/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 10596020,
+  "number_of_tokens": 278091459,
+  "min_length_tokens": 2,
+  "max_length_tokens": 20423,
+  "number_of_characters": 807674760,
+  "min_length_characters": 1,
+  "max_length_characters": 68505
+}

data/flashback-ovrigt/flashback-ovrigt.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Other'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Other
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Other subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Other subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 10.60M
+- **Number of tokens (Llama 3)**: 278.09M
+- **Average document length in tokens (min, max)**: 26.24 (2, 20.42K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-ovrigt_00000001",
+  "text": "Så slutligen ytterligare en fråga, hur ska han få reda på om det är sant?",
+  "source": "flashback-ovrigt",
+  "added": "2026-07-19",
+  "created": "2006-07-31, 2006-07-31",
+  "token_count": 27
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Other*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Other (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-ovrigt/flashback-ovrigt.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b5d6cea5a6975314828c6f8a2577a9284dcc3dc2e3aba5ee1a97bf43b21f362
+size 668465348

data/flashback-politik/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-politik` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-politik"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-politik.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-politik.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-politik/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 49832288,
+  "number_of_tokens": 1490319132,
+  "min_length_tokens": 2,
+  "max_length_tokens": 65308,
+  "number_of_characters": 4458096811,
+  "min_length_characters": 1,
+  "max_length_characters": 130613
+}

data/flashback-politik/flashback-politik.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Politics'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Politics
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Politics subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Politics subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 49.83M
+- **Number of tokens (Llama 3)**: 1.49B
+- **Average document length in tokens (min, max)**: 29.91 (2, 65.31K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-politik_00000001",
+  "text": "Sedan begriper jag förstås att rekreationsbruk av cannabis inte är missbruk eller självmedicinering.",
+  "source": "flashback-politik",
+  "added": "2026-05-30",
+  "created": "2018-10-28, 2018-10-28",
+  "token_count": 33
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Politics*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Politics (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-politik/flashback-politik.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20394f8cd43189a573913a148cb22dac247687a6b42674996d9afdd9f0aaaf1f
+size 3567073575

data/flashback-samhalle/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-samhalle` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-samhalle"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-samhalle.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-samhalle.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-samhalle/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 49485677,
+  "number_of_tokens": 1446029177,
+  "min_length_tokens": 2,
+  "max_length_tokens": 4500,
+  "number_of_characters": 4244469864,
+  "min_length_characters": 1,
+  "max_length_characters": 13920
+}

data/flashback-samhalle/flashback-samhalle.md

@@ -0,0 +1,95 @@
+---
+pretty_name: 'Flashback: Society'
+language:
+- sv
+license: cc-by-4.0
+license_name: CC-BY 4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- text-generation
+- fill-mask
+task_ids:
+- language-modeling
+source_datasets:
+- original
+domains:
+- Social Media
+---
+
+# Dataset Card for Flashback: Society
+
+<!-- START-SHORT DESCRIPTION -->
+Sentences from the Society subforum of [Flashback](https://www.flashback.org), a large Swedish internet forum.
+<!-- END-SHORT DESCRIPTION -->
+
+This subset contains sentences from the Society subforum of Flashback, one of Sweden's largest internet forums. Each sample corresponds to a single sentence — the source corpus supplies sentences in scrambled order for copyright and privacy reasons, so the original post context is not recoverable.
+
+## Dataset Description
+
+<!-- START-DESC-STATS -->
+- **Number of samples**: 49.49M
+- **Number of tokens (Llama 3)**: 1.45B
+- **Average document length in tokens (min, max)**: 29.22 (2, 4.50K)
+<!-- END-DESC-STATS -->
+
+
+## Dataset Structure
+An example from the dataset looks as follows.
+
+<!-- START-SAMPLE -->
+```py
+{
+  "id": "flashback-samhalle_00000001",
+  "text": "och högskoleutbildning.",
+  "source": "flashback-samhalle",
+  "added": "2026-05-30",
+  "created": "2007-10-01, 2007-10-01",
+  "token_count": 10
+}
+```
+
+### Data Fields
+
+An entry in the dataset consists of the following fields:
+
+- `id` (`str`): A unique identifier for each document.
+- `text` (`str`): The content of the document.
+- `source` (`str`): The source of the document (see [Source Data](#source-data)).
+- `added` (`str`): The date when the document was added to this collection.
+- `created` (`str`): The date range when the document was originally created.
+- `token_count` (`int`): The number of tokens in the sample computed using the Llama 3 tokenizer.
+<!-- END-SAMPLE -->
+
+### Dataset Statistics
+
+<!-- START-DATASET PLOTS -->
+<p align="center">
+<img src="./images/dist_document_length.png" width="600" style="margin-right: 10px;" />
+</p>
+<!-- END-DATASET PLOTS -->
+
+
+## Additional Information
+
+### Limitations
+
+Each sample is a single sentence. The source corpus deliberately scrambles sentence order for copyright and privacy protection, so the original post context is not recoverable.
+
+The source material is tokenized and does not preserve the original spacing between tokens. We approximate readable spacing when reconstructing plain text.
+
+As a forum corpus, this material contains informal, colloquial, and potentially offensive language.
+
+### Source
+
+We download the corpus from Språkbanken Text's official source file for *Flashback: Society*.
+
+### Citation Information
+
+If you use this subset, cite the original source release:
+
+> Språkbanken Text (2025). Flashback: Society (updated: 2025-04-01). [Data set]. Språkbanken Text.
+
+## License Information
+
+The source corpus is distributed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

data/flashback-samhalle/flashback-samhalle.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c59e2f5f05482ec5bfeb182ba22f5aeb7289190a01bb866af3b8ce99e48ca742
+size 3424365069

data/flashback-sport/create.py

@@ -0,0 +1,134 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "datasets>=3.0.0",
+#     "dynaword",
+#     "pydantic>=2.10.4",
+#     "transformers>=4.47.1",
+# ]
+#
+# [tool.uv.sources]
+# dynaword = { path = "../.." }
+# ///
+
+"""Download and process Språkbanken Text's `flashback-sport` corpus."""
+
+import bz2
+import logging
+import urllib.request
+import xml.etree.ElementTree as ET
+from datetime import date
+from pathlib import Path
+
+import datasets
+from datasets import Dataset
+from dynaword.process_dataset import (
+    add_token_count,
+    ensure_column_order,
+    remove_duplicate_text,
+    remove_empty_texts,
+)
+
+logger = logging.getLogger(__name__)
+
+SOURCE = "flashback-sport"
+XML_URL = "https://spraakbanken.gu.se/resurser/meningsmangder/flashback-sport.xml.bz2"
+CACHE_FILE = Path(__file__).parent / "cache" / "flashback-sport.xml.bz2"
+
+NO_SPACE_BEFORE = (".", ",", ";", ":", "!", "?", ")", "]", "}", "»")
+NO_SPACE_AFTER = ("(", "[", "{", "«")
+
+
+def download_if_needed(url: str, path: Path) -> None:
+    if path.exists():
+        logger.info("Using cached %s", path)
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    logger.info("Downloading %s to %s", url, path)
+    urllib.request.urlretrieve(url, path)
+    logger.info("Download complete")
+
+
+def format_date(value: str) -> str:
+    return f"{value[:4]}-{value[4:6]}-{value[6:8]}"
+
+
+def detokenize(tokens: list[str]) -> str:
+    text = " ".join(t.strip() for t in tokens if t and t.strip())
+    for token in NO_SPACE_BEFORE:
+        text = text.replace(f" {token}", token)
+    for token in NO_SPACE_AFTER:
+        text = text.replace(f"{token} ", token)
+    return " ".join(text.split()).strip()
+
+
+def sentence_text(sentence: ET.Element) -> str:
+    tokens = [(t.text or "") for t in sentence.iter("token")]
+    return detokenize(tokens)
+
+
+def iter_sentences(bz2_path: Path):
+    current_created = ""
+    index = 0
+    today = date.today().strftime("%Y-%m-%d")
+    with bz2.open(bz2_path, "rb") as f:
+        parent_stack: list[ET.Element] = []
+        for event, elem in ET.iterparse(f, events=("start", "end")):
+            if event == "start":
+                parent_stack.append(elem)
+                if elem.tag == "text":
+                    datefrom = elem.attrib.get("datefrom", "")
+                    dateto = elem.attrib.get("dateto", "")
+                    current_created = (
+                        f"{format_date(datefrom)}, {format_date(dateto)}"
+                        if datefrom and dateto
+                        else ""
+                    )
+            else:
+                if elem.tag == "sentence":
+                    text = sentence_text(elem)
+                    if text:
+                        index += 1
+                        yield {
+                            "id": f"{SOURCE}_{index:08d}",
+                            "text": text,
+                            "source": SOURCE,
+                            "added": today,
+                            "created": current_created,
+                        }
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                elif elem.tag in ("text", "paragraph"):
+                    if len(parent_stack) >= 2:
+                        parent_stack[-2].remove(elem)
+                    elem.clear()
+                parent_stack.pop()
+
+
+def build_dataset(bz2_path: Path) -> Dataset:
+    datasets.disable_caching()
+    dataset = Dataset.from_generator(lambda: iter_sentences(bz2_path))
+    dataset = remove_empty_texts(dataset)
+    dataset = remove_duplicate_text(dataset)
+    dataset = add_token_count(dataset)
+    dataset = ensure_column_order(dataset)
+    return dataset
+
+
+def main() -> None:
+    download_if_needed(XML_URL, CACHE_FILE)
+    dataset = build_dataset(CACHE_FILE)
+    output_path = Path(__file__).parent / f"{SOURCE}.parquet"
+    logger.info("Writing %s rows to %s", len(dataset), output_path)
+    dataset.to_parquet(output_path)
+
+
+if __name__ == "__main__":
+    log_path = Path(__file__).parent / f"{SOURCE}.log"
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        handlers=[logging.StreamHandler(), logging.FileHandler(log_path)],
+    )
+    main()

data/flashback-sport/descriptive_stats.json

@@ -0,0 +1,9 @@
+{
+  "number_of_samples": 18148623,
+  "number_of_tokens": 470500306,
+  "min_length_tokens": 2,
+  "max_length_tokens": 10365,
+  "number_of_characters": 1346398365,
+  "min_length_characters": 1,
+  "max_length_characters": 20327
+}