imliuyu/qwen3.5-9b-dsr1-cot-sft
This model is a fine-tuned variant of Qwen3.5-9B, built on top of the original model and trained using a carefully selected subset of STEM and math data from the DeepSeek-R1 dataset. It is optimized for more concise and rational chain-of-thought reasoning.
中文 English
Model Highlights
Below is a comparison of three models:
- sft_org: The official original version, Qwen3.5-9B
- sft_claude: Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled
- sft_r1: This model
Key Evaluation Metrics
(During testing, max_tokens was set to 8192; outputs exceeding this limit were truncated.)
Core Advantages
1. Think/Output Token Ratio: Both sftclaude and sftr1 (chain-of-thought fine-tuned models) significantly reduce the number of reasoning tokens compared to the original sft_org. This model achieves an even greater reduction.
2. Accuracy (green area in the right chart): When excluding truncation effects, both sftclaude and sftr1 show some accuracy drop after chain-of-thought tuning. However, this model only experiences a minimal decline of 1.1%.
3. Pass Rate (blue area in the right chart): In real-world applications, cost factors such as token usage and inference time must be considered. After accounting for truncation costs, both sftclaude and sftr1 demonstrate significant improvements in pass rate, with this model achieving the most notable gain.
Deployment & Usage
This model is fully compatible with the original Qwen3.5 in terms of deployment and usage. Please refer directly to the official quickstart guide.
Example:
vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3Note: In vLLM version 0.15.x, the think content may be incorrectly concatenated into the content field. In this case, you can use:
--reasoning-parser deepseek_r1Additional Information
For more details on fine-tuning, evaluation procedures, datasets, and code, please refer to our GitHub repository: https://github.com/codingmiao/qwen35-short-cot