From 102c50e4d6bd4d8c874ca737c8312bc2cd691bea Mon Sep 17 00:00:00 2001 From: ai-modelscope Date: Tue, 22 Apr 2025 08:58:53 +0800 Subject: [PATCH] Update README.md --- README.md | 66 ++++++++++++++++---------------- generation_config.json | 14 ------- model-00001-of-00004.safetensors | 3 -- model-00002-of-00004.safetensors | 3 -- model-00003-of-00004.safetensors | 3 -- model-00004-of-00004.safetensors | 3 -- 6 files changed, 34 insertions(+), 58 deletions(-) delete mode 100644 generation_config.json delete mode 100644 model-00001-of-00004.safetensors delete mode 100644 model-00002-of-00004.safetensors delete mode 100644 model-00003-of-00004.safetensors delete mode 100644 model-00004-of-00004.safetensors diff --git a/README.md b/README.md index 311122c..5556998 100644 --- a/README.md +++ b/README.md @@ -4,52 +4,53 @@ license: apache-2.0 ## 📖 Introduction -# DistilQwen2.5-DS3-0324 系列快思考推理模型 +# DistilQwen2.5-DS3-0324 Series: Fast-Thinking Reasoning Models -## 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 +## Overview +In response to the industry challenge of balancing efficient reasoning with cognitive capabilities, the DistilQwen2.5-DS3-0324 series innovatively transfers the fast-thinking capabilities of DeepSeekV3-0324 to lightweight models. Through a two-stage distillation framework, this series achieves high performance while delivering: +- **Enhanced Reasoning Speed**: Reduces output tokens by 60-80% (compared to slow-thinking models) +- **Reduced Resource Consumption**: Suitable for edge computing deployment +- **Elimination of Cognitive Bias**: Proprietary trajectory alignment technology -## 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 +## Core Innovations +### 1. Fast-Thinking Distillation Framework +- **Stage 1: Fast-Thinking CoT Data Collection** + - **Long-to-Short Rewriting**: Extracts key reasoning steps from DeepSeek-R1 + - **Teacher Model Distillation**: Captures the rapid reasoning trajectories of DeepSeekV3-0324 -- **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 +- **Stage 2: CoT Trajectory Cognitive Alignment** + - **Dynamic Difficulty Grading** (Easy/Medium/Hard) + - LLM-as-a-Judge evaluates small model comprehensibility + - Simple chain expansion → Adds necessary steps + - Hard chain simplification → Removes high-level logical leaps + - **Validation Mechanism**: Iterative optimization until all data reaches "Medium" rating -### 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） +### 2. Performance Breakthroughs +- **32B Model** approaches the performance of closed-source models with 10x the parameters on the GPQA Diamond benchmark +- **Significant Improvement in Reasoning Efficiency** (see comparison table below) -| 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 | +| Model | MMLU_PRO Tokens | AIME2024 Tokens | Speed Gain | +|--------------------------------|-----------------|-----------------|------------| +| DistilQwen2.5-R1-32B (Slow-Thinking) | 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 +## Technical Advantages +- **Two-Stage Distillation**: First compresses reasoning length, then aligns cognitive trajectories +- **Dynamic Data Optimization**: Adaptive difficulty adjustment ensures knowledge transferability +- **Open-Source Compatibility**: Fine-tuned based on the Qwen2.5 base model -## 🚀 快速开始 +## 🚀 Quick Start ```python -from modelscope import AutoModelForCausalLM, AutoTokenizer +from transformers import AutoModelForCausalLM, AutoTokenizer +device = "cuda" # the device to load the model onto model = AutoModelForCausalLM.from_pretrained( - "PAI/DistilQwen2.5-DS3-0324-7B", + "alibaba-pai/DistilQwen2.5-DS3-0324-7B", torch_dtype="auto", device_map="auto" ) -tokenizer = AutoTokenizer.from_pretrained("PAI/DistilQwen2.5-DS3-0324-7B") +tokenizer = AutoTokenizer.from_pretrained("alibaba-pai/DistilQwen2.5-DS3-0324-7B") prompt = "Give me a short introduction to large language model." messages=[ @@ -72,4 +73,5 @@ generated_ids = [ ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] + ``` \ No newline at end of file diff --git a/generation_config.json b/generation_config.json deleted file mode 100644 index a753841..0000000 --- a/generation_config.json +++ /dev/null @@ -1,14 +0,0 @@ -{ - "bos_token_id": 151643, - "do_sample": true, - "eos_token_id": [ - 151645, - 151643 - ], - "pad_token_id": 151643, - "repetition_penalty": 1.05, - "temperature": 0.7, - "top_k": 20, - "top_p": 0.8, - "transformers_version": "4.46.1" -} diff --git a/model-00001-of-00004.safetensors b/model-00001-of-00004.safetensors deleted file mode 100644 index a58471d..0000000 --- a/model-00001-of-00004.safetensors +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:2936da2eb6666f60def42e354b4e36a9bad19ebdf5090cbc02c0834f92c27df5 -size 4877660776 diff --git a/model-00002-of-00004.safetensors b/model-00002-of-00004.safetensors deleted file mode 100644 index 7a2f0f7..0000000 --- a/model-00002-of-00004.safetensors +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:11aab5fc3313e10f1aaa9438a2d74cfe5c59538dd8e6506ec216926887cc0820 -size 4932751008 diff --git a/model-00003-of-00004.safetensors b/model-00003-of-00004.safetensors deleted file mode 100644 index a22514e..0000000 --- a/model-00003-of-00004.safetensors +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:2053eee73faf67773f7e8e042b2f3e9bf599edb6bca4e77a2cd0d83d375852de -size 4330865200 diff --git a/model-00004-of-00004.safetensors b/model-00004-of-00004.safetensors deleted file mode 100644 index bcdc21c..0000000 --- a/model-00004-of-00004.safetensors +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:30be9d72da5b630c956bcd0fa3725246d1c5891df7f9d0bec00f65ea7145114a -size 1089994880