---
title: qwen3.5-9b-dsr1-cot-sft
canonical_url: "https://www.modelscope.cn/models/codingmiao/qwen3.5-9b-dsr1-cot-sft"
md_url: "https://www.modelscope.cn/models/codingmiao/qwen3.5-9b-dsr1-cot-sft.md"
repository: codingmiao/qwen3.5-9b-dsr1-cot-sft
chinese_name: qwen3.5-9b-dsr1-cot-sft
last_updated: 2026-03-27
license: "Apache License 2.0"
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3_5
architectures:
  - Qwen3_5ForConditionalGeneration
base_model_relation: finetune
parameters: 9.4B
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - PyTorch
language:
  - zh
  - en
downloads: 95
stars: 1
tags:
  - cot
---

# qwen3.5-9b-dsr1-cot-sft

> qwen3.5-9b-dsr1-cot-sft - codingmiao 在 ModelScope 开源的模型。本模型是在Qwen3.5-9B基础上，从 deepseek-r1 数据集中精选部分stem、math类数据微调其思维链的版本，拥有更为简洁、理性的思考能力。

codingmiao/qwen3.5-9b-dsr1-cot-sft 是 ModelScope 魔搭社区上的 9.4B 参数text-generation模型，采用 Apache License 2.0 许可。

- **Repository**: codingmiao/qwen3.5-9b-dsr1-cot-sft
- **License**: Apache License 2.0
- **Tasks**: text-generation
- **Parameters**: 9.4B
- **Tags**: cot
- **Downloads**: 95
- **Stars**: 1
- **Last updated**: 2026-03-27

Source: https://www.modelscope.cn/models/codingmiao/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](README_zh.md) [English](README.md)


# Model Highlights

Below is a comparison of three models:

* **sft_org**: The official original version, [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)
* **sft_claude**: [Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled](https://huggingface.co/Jackrong/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**.)*

| Model      | Truncation Rate | Overall Accuracy | **True Accuracy** | Avg Tokens | Compression Ratio |
| ---------- | --------------- | ---------------- | ----------------- | ---------- | ----------------- |
| sft_org    | 82.1%           | 28.8%            | **95.0%**         | 7753       | 1.0x              |
| sft_claude | 30.1%           | 60.0%            | 82.3%             | 4141       | 1.87x             |
| **sft_r1** | **22.4% ↓**     | **73.3% ↑**      | 93.9%             | **3071 ↓** | **2.52x ↓**       |

---

## Core Advantages

![Token Distribution](imgs/06_token_distribution.png)

> **1. Think/Output Token Ratio**:
> Both sft_claude and sft_r1 (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.

---

![Accuracy Analysis](imgs/05_truncation_impact.png)

> **2. Accuracy** *(green area in the right chart)*:
> When excluding truncation effects, both sft_claude and sft_r1 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 sft_claude and sft_r1 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](https://huggingface.co/Qwen/Qwen3.5-9B#quickstart).

Example:

```
vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
```

**Note:**
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_r1
```

---

# Additional 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](https://github.com/codingmiao/qwen35-short-cot)