codepause-phase7-qwen25-coder-7b-gguf

CodePause Phase 7 — Qwen2.5-Coder-7B ThinkAnywhere GGUF

This is a GGUF export of the CodePause Phase 7 model: a Qwen2.5-Coder-7B-Instruct based model fine-tuned with QLoRA for structured code reasoning using <thinkanywhere> blocks.

Intended use

Use in LM Studio or llama.cpp-compatible runtimes for code generation experiments.

Recommended prompt style:

You MUST answer using exactly this format:

<thinkanywhere>
Briefly explain the algorithm, edge cases, and complexity.
</thinkanywhere>

```python
# final code only

Task: Write a Python function longest_unique_substring(s).

## Training summary

- Base model: `Qwen/Qwen2.5-Coder-7B-Instruct`
- Method: QLoRA 4-bit NF4
- Dataset: CodePause Dataset v7
- Dataset size: 150 examples
- Mix: 70% examples with structured reasoning, 30% plain code
- Epochs: 3
- Final artifact: F16 GGUF

## Known limitations

- The model can generate correct code, but `<thinkanywhere>` tag adherence may still require strong prompt formatting.
- This F16 GGUF is large (~15.2GB). Quantized Q4_K_M export is recommended for faster local inference.

## Local loading

Load the `.gguf` file in LM Studio using a Qwen/ChatML-compatible prompt template.