Update README.md

README.md

@@ -31,26 +31,17 @@ configs:
 
 # Classical-Swedish Allusions: Gold Benchmark
 
-A small, carefully curated reference set of cross-lingual allusions from classical Greek and Latin into Swedish literature. Used as a held-out evaluation set for the broader [`Ericu950/classical-swedish-citations-v2`](https://huggingface.co/datasets/Ericu950/classical-swedish-citations-v2) project.
+A small curated reference set of cross-lingual allusions from classical Greek and Latin into Swedish literature. Used as a held-out evaluation set for the [`Ericu950/classical-swedish-citations-v2`](https://huggingface.co/datasets/Ericu950/classical-swedish-citations-v2) project.
 
 ## Purpose
 
-This is **not training data**. It is a fixed, public reference standard. Any cross-lingual allusion-detection system should aspire to recover these. The set was archived publicly **before** evaluating any system against it, so reported recall scores are honest.
+This is **not training data**. It is a fixed, public reference standard. Our cross-lingual allusion-detection system will aspire to recover these. The set was archived publicly **before** evaluating any system against it.
 
 The pairs were chosen to span several difficulty modes:
 
-- **Verbatim Latin quotations** embedded in Swedish text
-- **Translated proverbial tags** (e.g. *parturient montes* → "skohla bärgen föda")
-- **Homeric formulas** rendered into Swedish ("ἔπεα πτερόεντα" → "de bevingade orden")
-- **Biblical and Pauline phrases** echoed in 19th-c. Swedish prose
-- **Extended paraphrases** of classical passages
-- **Stylistic imitations** of classical forms (epinician, Homeric simile, etc.)
-
-Many of these are **outside the reach of current retrieval-based systems**. That is intentional. The benchmark defines the open problem.
-
 ## Files
 
-- `manual.tsv` — the raw gold pairs, tab-separated, headless, two columns: source_sentence, target_sentence. The source side matches sentences as they appear in the Perseus Digital Library JSONL release; the target side matches sentences after SaT-12l-sm segmentation of the corresponding Swedish text.
+- `manual.tsv` — the raw gold pairs, tab-separated, headless, two columns: source_sentence, target_sentence. The source side matches sentences as they appear in the Perseus Digital Library, sentence tokenized with SaT-12l-sm, and  Swedish text tokenized with the same model
 
 - `gold_benchmark.jsonl` — the expanded benchmark, one record per (source_window, target_window) positional variant. Each gold pair appears multiple times: the focal sentence can sit at positions 0..4 within its 5-sentence window on each side. This tests positional robustness.
 
@@ -75,35 +66,6 @@ Many of these are **outside the reach of current retrieval-based systems**. That
 | `target_book` | str | Swedish book identifier |
 | `target_sent_idx` | int | Sentence index in the SaT-segmented Swedish text |
 
-A single source allusion may have multiple `gold_pair_id` realizations if the same Greek/Latin sentence appears in multiple Perseus files (e.g. Vergil's *auri sacra fames* appearing in *Aeneid* and again in Quintilian's quotation).
-
-## Evaluation protocol
-
-The recommended evaluation metric is **gold_pair_id-level recall**: a system succeeds on a gold pair if it recovers **any** of that pair's (source_window, target_window) records. Positional variants exist to test robustness, not to inflate the count.
-
-```python
-from datasets import load_dataset
-import json
-
-gold = load_dataset("Ericu950/classical-swedish-allusions-gold", split="test")
-gold_pair_ids = set(rec["gold_pair_id"] for rec in gold)
-
-# After running your pipeline against the source/target corpora:
-recovered_pair_ids = set()  # populate from your pipeline output
-
-recall = len(recovered_pair_ids & gold_pair_ids) / len(gold_pair_ids)
-print(f"Gold recall: {recall:.2%}")
-```
-
-## Reference baseline
-
-The pipeline described in [`Ericu950/classical-swedish-citations-v2`](https://huggingface.co/datasets/Ericu950/classical-swedish-citations-v2) recovers approximately 3 of these gold pairs at the time of the benchmark's release. The remaining ~12 require capabilities the current architecture lacks:
-
-- Recognizing **translated proverbial tags** where surface forms diverge entirely (Latin→Swedish translation memory)
-- Recognizing **stylistic imitations** as kindred rather than as citations
-- Recognizing **biblical thematic echoes** without explicit lexical overlap
-
-The benchmark is intended to define these as open problems, not to suggest the current pipeline is failing.
 
 ## Languages
 
@@ -113,17 +75,14 @@ The benchmark is intended to define these as open problems, not to suggest the c
 ## Source corpora
 
 - Greek + Latin: Perseus Digital Library, sentence-segmented
-- Swedish: hand-curated texts under `00_texts/` in the pipeline repo, SaT-12l-sm segmented
+- Swedish: hand-curated texts by Carl Magnus Juliusson and Eric Cullhed
 
 ## License
 
 CC BY 4.0. Free to use for any purpose with attribution.
 
-## Citation
-
-If you use this benchmark, please cite the accompanying paper (forthcoming) and this dataset.
 
 ## Companion resources
 
-- Production pipeline output (~3,000 confirmed citations): [`Ericu950/classical-swedish-citations-v2`](https://huggingface.co/datasets/Ericu950/classical-swedish-citations-v2)
+- Production pipeline output (these samples are not in Litteraturbanken) (~3,000 confirmed citations): [`Ericu950/classical-swedish-citations-v2`](https://huggingface.co/datasets/Ericu950/classical-swedish-citations-v2)
 - Bi-encoder models: [`Ericu950/intertext-classical-swedish-sentence`](https://huggingface.co/Ericu950/intertext-classical-swedish-sentence), [`Ericu950/intertext-classical-swedish-window`](https://huggingface.co/Ericu950/intertext-classical-swedish-window)