Upload README.md with huggingface_hub

README.md

@@ -1,7 +1,7 @@
 ---
 language:
 - en
-license: apache-2.0
+license: mit
 base_model: Qwen/Qwen2.5-Coder-7B-Instruct
 tags:
 - qwen2.5
@@ -9,26 +9,27 @@ tags:
 - code
 - cli
 - debugging
+- developer-tools
 - lora
+- mlx
+- gguf
+- ollama
 model-index:
-- name: YOLO-7B-Qwen-Coder
+- name: YOLO-Coder-8B
   results: []
-library_name: mlx
 ---
 
-<center><img src="https://cdn-uploads.huggingface.co/production/uploads/69518ec2b519522f5659da57/tEU3vljyhBNa5PTxPwUWa.png" width="500"></center>
+# YOLO-Coder-8B
 
-# YOLO-7B-Qwen-Coder
+A fine-tuned version of [Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) specialized in diagnosing CLI errors and generating a single, precise bash fix command. Runs on any machine with 6GB free RAM. Responds in 1–3s on Apple Silicon.
 
-A fine-tuned version of [Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) specialized in diagnosing CLI errors and generating a single, precise bash fix command.
-
-Part of the [yolo-coder](https://github.com/erdemozkan/yolo-coder) project — an automated CLI repair tool that wraps any command, catches failures, and fixes them locally using a local LLM.
+The **high-accuracy sibling** of [YOLO-Coder-1.5B](https://huggingface.co/erdemozkan/YOLO-Coder-1.5B). Part of the [YOLO-CODER](https://github.com/erdemozkan/YOLO-CODER) project — an automated CLI repair tool that wraps any command, catches failures, and fixes them locally using a three-layer pipeline: deterministic interceptors → fix memory → local LLM.
 
 ---
 
 ## What it does
 
-Given a CLI error message and surrounding code context, the model outputs **exactly one bare bash command** to fix the problem. No explanation. No markdown. No backticks. Just the fix.
+Given a CLI error message, the model outputs **exactly one bare bash command** to fix the problem. No explanation. No markdown. No backticks. Just the fix.
 
 ```
 Input:  ModuleNotFoundError: No module named 'requests'
@@ -36,12 +37,44 @@ Output: pip install requests
 ```
 
 ```
-Input:  IndexError: list index out of range  (line 7: print(items[2]))
-Output: python3 yoco_replace.py app.py 7 'print(items[1])'
+Input:  PermissionError: [Errno 13] Permission denied: '/usr/local/bin/tool'
+Output: sudo chmod +x /usr/local/bin/tool
+```
+
+```
+Input:  npm ERR! code EACCES
+        npm ERR! syscall mkdir
+        npm ERR! path /usr/local/lib/node_modules
+Output: sudo npm install -g typescript
 ```
 
 ---
 
+## Benchmark Results (YOLO-Bench)
+
+Evaluated on **218 verified CLI error examples** across 15 categories using structural match scoring (flag-order-independent, compound-command-aware).
+
+| Mode | Structural Match |
+|---|---|
+| Raw LLM (no pipeline) | **59.2%** |
+| Pipeline × 1 (interceptors + LLM) | **72.0%** |
+| Pipeline × 3 (interceptors + memory + 3 LLM attempts) | **77.1%** |
+
+**vs. baselines (raw LLM only):**
+
+| Model | Structural Match |
+|---|---|
+| **YOLO-Coder-8B** | **59.2%** |
+| YOLO-Coder-1.5B | 66.5% |
+| GPT-4o | 48.6% |
+| Claude Sonnet 4.6 | 60.1% |
+
+> YOLO-Coder-8B pipeline×3 hits 77.1% — the highest of any model tested, including GPT-4o and Claude Sonnet.
+
+Dataset and scoring code: [github.com/erdemozkan/YOLO-CODER/tree/main/benchmark](https://github.com/erdemozkan/YOLO-CODER/tree/main/benchmark)
+
+---
+
 ## Model Details
 
 | Property | Value |
@@ -51,34 +84,24 @@ Output: python3 yoco_replace.py app.py 7 'print(items[1])'
 | LoRA rank | 8 |
 | LoRA scale | 20.0 |
 | Layers trained | 28 |
-| Training iterations | 800 |
-| Learning rate | 5e-6 |
-| Batch size | 2 (grad accumulation × 4 = effective 8) |
-| Max sequence length | 1024 |
-| Training hardware | Apple Silicon M-series (16GB) |
-| Final train loss | 0.217 |
-| Final val loss | 0.260 |
+| Training iterations | 500 |
+| Learning rate | 1e-5 |
+| Batch size | 2 |
+| Max sequence length | 2048 |
+| Training hardware | Apple Silicon M-series |
+| Training examples | 6,719 |
+| Model size (GGUF Q4_K_M) | ~4.4GB |
+| RAM required | ~6GB |
 
 ---
 
 ## Training Data
 
-Trained on a custom dataset of **2,250 CLI error/fix pairs** covering:
+Trained on **6,719 CLI error → fix pairs** across 15 categories:
 
-- Python: `SyntaxError`, `ImportError`, `ModuleNotFoundError`, `AttributeError`, `IndexError`, `KeyError`, `TypeError`, `ZeroDivisionError`, `PermissionError`, `FileNotFoundError`
-- pip: missing packages, `--break-system-packages`, hash mismatches, permissions
-- Node.js: `Cannot find module`, `MODULE_NOT_FOUND`
-- npm: `ENOENT`, `ERESOLVE`, `EACCES`
-- TypeScript: `TS2304`, `TS2339`
-- Docker: image not found, port conflicts, container name collisions, daemon not running
-- Git: merge conflicts, detached HEAD, push rejected, not a repo
-- Web frameworks: Next.js, FastAPI, Flask, Express
-- Auth errors: JWT, OAuth, session issues
-- Async/CORS errors
+`Python` · `pip` · `Node.js` · `npm` · `TypeScript` · `Docker` · `Git` · `shell` · `Cargo/Rust` · `SSH` · `database` · `venv/conda` · `make/cmake` · `cloud (AWS/GCP)` · `yarn`
 
-Format: ChatML with a strict system prompt enforcing single-command output.
-
-90/10 train/validation split → 2025 train, 225 validation examples.
+Format: ChatML with a system prompt enforcing single-command output. Each example is a real error a developer hit in a real terminal.
 
 ---
 
@@ -86,63 +109,38 @@ Format: ChatML with a strict system prompt enforcing single-command output.
 
 | File | Description |
 |---|---|
-| `YOLO-7B-Qwen-q4.gguf` | Q4_K_M quantized GGUF (~4.4GB) — recommended for Ollama |
-| `YOLO-7B-Qwen-finetuned.gguf` | f16 GGUF (~14GB) — full quality, for requantization |
+| `YOLO-Coder-8B-Q4_K_M.gguf` | Q4_K_M quantized GGUF (~4.4GB) — recommended for Ollama |
 | `safetensors/` | fp16 HuggingFace safetensors — for further fine-tuning |
-| `adapter/` | Raw LoRA adapter weights (MLX format) |
 
 ---
 
-## Usage with Ollama
+## Quickstart with Ollama
 
 ```bash
-# Download the Modelfile
-curl -O https://raw.githubusercontent.com/erdemozkan/yolo-coder/main/YOLO-MODEL-FILES/Modelfile-7B
-
-# Pull the GGUF and register
-ollama create yolo-7b -f Modelfile-7B
+# Pull and run directly
+ollama run hf.co/erdemozkan/YOLO-Coder-8B
 
 # Test it
-ollama run yolo-7b "ModuleNotFoundError: No module named 'flask'"
+ollama run hf.co/erdemozkan/YOLO-Coder-8B "ModuleNotFoundError: No module named 'flask'"
 # → pip install flask
 ```
 
 ---
 
-## Usage with yolo-coder
+## Usage with YOLO-CODER
 
 ```bash
-git clone https://github.com/erdemozkan/yolo-coder
-cd yolo-coder
-pip install -e .
-
-# Use the 7B model
-yoco --model yolo-7b python3 myapp.py
-```
-
----
-
-## Usage with Transformers
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+pip install yolo-coder
 
-model_id = "erdemozkan/YOLO-7B-Qwen-Coder"
-tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="safetensors")
-model = AutoModelForCausalLM.from_pretrained(model_id, subfolder="safetensors")
+# Use with the 8B model (default)
+yoco python3 myapp.py
 
-messages = [
-    {"role": "system", "content": "You are a CLI repair tool. Output ONLY a single bare bash command to fix the error. No explanation. No markdown. No backticks."},
-    {"role": "user", "content": "ModuleNotFoundError: No module named 'requests'"}
-]
-
-text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-inputs = tokenizer(text, return_tensors="pt")
-outputs = model.generate(**inputs, max_new_tokens=64, temperature=0.1)
-print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
-# → pip install requests
+# Or explicitly
+yoco --model hf.co/erdemozkan/YOLO-Coder-8B python3 myapp.py
 ```
 
+YOLO-CODER adds the full pipeline on top: 91 deterministic interceptors fire first (<1ms), then fix memory, then the LLM as a last resort. This pushes the effective fix rate to 77.1%.
+
 ---
 
 ## Prompt Format (ChatML)
@@ -151,21 +149,39 @@ print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_t
 <|im_start|>system
 You are a CLI repair tool. Output ONLY a single bare bash command to fix the error. No explanation. No markdown. No backticks.<|im_end|>
 <|im_start|>user
-{error message}
-<|im_end|>
+[Linux] $ python3 myapp.py
+Error:
+ModuleNotFoundError: No module named 'requests'
+FIX:<|im_end|>
 <|im_start|>assistant
+pip install requests<|im_end|>
 ```
 
 ---
 
+## 1.5B vs 8B
+
+| | YOLO-Coder-1.5B | YOLO-Coder-8B |
+|---|---|---|
+| Size | ~941MB | ~4.4GB |
+| RAM needed | ~2GB | ~6GB |
+| Speed | <1s on Apple Silicon | 1–3s on Apple Silicon |
+| Raw accuracy | 66.5% | 59.2% |
+| Pipeline×3 accuracy | 71.1% | **77.1%** |
+| Best for | Speed, low-RAM machines | Hard errors, best accuracy |
+
+The 8B model gains the most on complex, multi-step, and rare errors where its larger context helps. The 1.5B is surprisingly strong on common errors and runs on almost anything.
+
+---
+
 ## Limitations
 
-- Outputs a single command only — not suitable for multi-step fixes without a wrapper
-- Trained on common CLI errors; rare or highly domain-specific errors will fall back to base model behavior
-- Not a general coding assistant — use base Qwen2.5-Coder for that
+- Single-command output only — not suitable for multi-step fixes without a wrapper
+- Complex or highly novel errors may fall through to suboptimal output
+- Not a general coding assistant
 
 ---
 
 ## License
 
-Apache 2.0 — same as the base model.
+MIT