---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3.5-2B-Base
---
# reasoning_summarizer

Fine-tuned text-only Qwen3.5 2B Base model for converting reasoning-chain text into JSON metadata.

## Format

Input is raw reasoning text. The model is trained to output only JSON:

```json
{"title":"...","sub_title":"...","summary":"...","cur_task":"..."}
```

No system prompt was used in training.

## Training

- Base model: `Qwen/Qwen3.5-2B-Base`
- Adapter source: `runs/qwen3_5_2b_reasoning_json_lora/best_adapter`
- LoRA rank: 32
- LoRA alpha: 64
- Training context length: 1024
- Text-only export: merged with `AutoModelForCausalLM`; no mmproj export

## Files

- `reasoning_summarizer_hf/`: merged Hugging Face safetensors checkpoint
- `reasoning_summarizer-f16.gguf`: F16 GGUF
- `reasoning_summarizer-Q4_K_M.gguf`: Q4 quantized GGUF
- `reasoning_summarizer-Q5_K_M.gguf`: Q5 quantized GGUF
- `reasoning_summarizer-Q6_K.gguf`: Q6 quantized GGUF
- `reasoning_summarizer-Q8_0.gguf`: Q8 quantized GGUF