75 lines
2.7 KiB
Markdown
75 lines
2.7 KiB
Markdown
---
|
||
license: apache-2.0
|
||
---
|
||
|
||
## 📖 Introduction
|
||
|
||
# DistilQwen2.5-DS3-0324 系列快思考推理模型
|
||
|
||
## 概述
|
||
在平衡高效推理与思维能力的行业挑战下，DistilQwen2.5-DS3-0324系列创新性地将DeepSeekV3-0324的快思考能力迁移到轻量模型中。通过两阶段蒸馏框架，该系列在保持高性能的同时实现：
|
||
- **推理速度提升**：输出token数减少60-80%（相比慢思考模型）
|
||
- **资源消耗降低**：适合边缘计算部署
|
||
- **认知偏差消除**：独创的轨迹对齐技术
|
||
|
||
## 核心创新
|
||
### 1. 快思考蒸馏框架
|
||
- **阶段一：快思考CoT数据收集**
|
||
- **Long-to-Short改写**：从DeepSeek-R1提炼关键推理步骤
|
||
- **教师模型蒸馏**：提取DeepSeekV3-0324的快速推理轨迹
|
||
|
||
- **阶段二：CoT轨迹认知对齐**
|
||
- **动态难度分级**（简单/中等/困难）
|
||
- LLM-as-a-Judge评估小模型可理解性
|
||
- 简单链扩展 → 补充必要步骤
|
||
- 困难链精简 → 移除高阶逻辑跳跃
|
||
- **验证机制**：迭代优化直至所有数据达"中等"评级
|
||
|
||
### 2. 性能突破
|
||
- **32B模型**在GPQA Diamond基准接近10倍参数量的闭源模型
|
||
- **推理效率**显著提升（见下表对比）
|
||
|
||
| 模型 | MMLU_PRO Tokens | AIME2024 Tokens | 速度增益 |
|
||
|--------------------------------|-----------------|-----------------|----------|
|
||
| DistilQwen2.5-R1-32B (慢思考) | 4198 | 12178 | 1x |
|
||
| DistilQwen2.5-DS3-0324-32B | 690 | 4177 | 5-8x |
|
||
|
||
## 技术优势
|
||
- **双阶段蒸馏**：先压缩推理长度，再对齐认知轨迹
|
||
- **动态数据优化**：自适应难度调整确保知识可迁移性
|
||
- **开源兼容**：基于Qwen2.5基座模型微调
|
||
|
||
## 🚀 快速开始
|
||
|
||
```python
|
||
from modelscope import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"PAI/DistilQwen2.5-DS3-0324-32B",
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained("PAI/DistilQwen2.5-DS3-0324-32B")
|
||
|
||
prompt = "Give me a short introduction to large language model."
|
||
messages=[
|
||
{"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant. You should think step-by-step."},
|
||
{"role": "user", "content": prompt},
|
||
]
|
||
text = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=False,
|
||
add_generation_prompt=True
|
||
)
|
||
model_inputs = tokenizer([text], return_tensors="pt").to(device)
|
||
|
||
generated_ids = model.generate(
|
||
model_inputs.input_ids,
|
||
max_new_tokens=2048，
|
||
)
|
||
generated_ids = [
|
||
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
|
||
]
|
||
|
||
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
||
``` |