38 lines
692 B
Markdown
38 lines
692 B
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- qwen3
|
||
|
|
- text-generation
|
||
|
|
- thinking-sft
|
||
|
|
- code
|
||
|
|
---
|
||
|
|
|
||
|
|
# Comm-C Qwen3-8B Thinking SFT
|
||
|
|
|
||
|
|
This checkpoint is a short 20-step SFT run based on Qwen3-8B.
|
||
|
|
|
||
|
|
Training target format:
|
||
|
|
|
||
|
|
```text
|
||
|
|
<think>
|
||
|
|
reasoning_content
|
||
|
|
</think>
|
||
|
|
final response
|
||
|
|
```
|
||
|
|
|
||
|
|
Training summary:
|
||
|
|
|
||
|
|
- Base model: Qwen3-8B
|
||
|
|
- Data: synthetic Comm-C distillation data from GLM-5.2
|
||
|
|
- Train rows: 4243
|
||
|
|
- Validation rows: 99
|
||
|
|
- Max length: 32768
|
||
|
|
- Global batch size: 64
|
||
|
|
- Learning rate: 2e-5
|
||
|
|
- Training steps: 20
|
||
|
|
- Output format: Hugging Face safetensors shards
|
||
|
|
|
||
|
|
This checkpoint is intended as a quick sanity-check artifact for thinking-format
|
||
|
|
SFT, not a final converged model.
|