贡献知识 → 融合进化 → 获取更强的全局模型
DeepFusion Protocol powered by View-Aligned Attention
Contribute Knowledge → Fuse & Evolve → Get Stronger Global Models
DeepFusion Protocol powered by View-Aligned Attention
三步完成知识融合与进化
Three steps to knowledge fusion and evolution
上传你的 AI 模型权重或梯度到蒸馏中心，支持 ONNX、PyTorch、SafeTensors 等格式
Upload your AI model weights or gradients to the distillation hub. Supports ONNX, PyTorch, SafeTensors and more
通过 View-Aligned Attention 机制，将异构模型的知识对齐并融合到全局 MoE
Using View-Aligned Attention mechanism to align and fuse knowledge from heterogeneous models into the Global MoE
下载改进后的全局模型，同时获得 贡献积分 奖励（Proof-of-Contribution）
Download the improved global model and earn Contribution Points (Proof-of-Contribution)
去中心化的 AI 知识融合协议
Decentralized AI Knowledge Fusion Protocol
帮助您把领域知识蒸馏到高效的 3B 本地模型
Distill your domain knowledge into an efficient 3B local model
上传您的 teacher 模型或知识数据（alpaca/QA/chat 格式）
Upload your teacher model or knowledge data (alpaca/QA/chat format)
使用 v7 蒸馏框架，自动生成 SFT 数据，训练 LoRA adapter
Use v7 framework to auto-generate SFT data and train LoRA adapter
在您的任务上评测效果，对比 Base 模型的提升
Evaluate on your custom tasks and compare with base model
下载优化后的 3B 模型，在本地 GPU/CPU 上部署使用
Download optimized 3B model and deploy on your GPU/CPU
14B Teacher → 3B Student，LoRA-MoE 知识蒸馏实测
14B Teacher → 3B Student, LoRA-MoE Knowledge Distillation Benchmarks
| 类别Category | Base 3B | DeepFusion v7 | 变化Delta |
|---|---|---|---|
| AI 与技术AI & Tech | 7.7 | 8.2 | +0.5 |
| 数学推理Math Reasoning | 6.4 | 6.9 | +0.5 |
| 代码生成Code Gen | 3.3 | 3.6 | +0.3 |
| 创意写作Creative Writing | 6.9 | 6.7 | -0.2 |
| 指令跟随Instruction Following | 5.2 | 4.9 | -0.3 |
| 常识问答Common Sense | 5.5 | 5.2 | -0.3 |
| 实用场景Practical | 7.8 | 7.4 | -0.4 |
| 逻辑推理Logic | 7.5 | 6.5 | -1.0 |
| 科学知识Science | 8.2 | 7.4 | -0.8 |
| 中文理解Chinese NLU | 7.8 | 6.7 | -1.1 |
| 总平均Overall | 6.63 | 6.35 | -0.28 |
* v7 改动：1500条SFT数据 + FM权重×6 + KL权重÷6（方法论蒸馏） * v7 changes: 1500 SFT samples + FM weight ×6 + KL weight ÷6 (methodology distillation)
| 版本Version | BETA (KL) | LoRA R | 数据量Data | 训练时间Time | 最佳提升Best Gain | 状态Status |
|---|---|---|---|---|---|---|
| v3 | 0.8 | 32 | 15K x 3ep | ~12h | - | KL 过高 KL too high |
| v4 | 0.3 | 64 | 30K x 1ep | 7.2h | Math +0.9 | 改进中 Improving |
| v5 (best) | 0.5 | 32 | 15K x 2ep | 6.5h | Science +0.7 | 当前最优 Current Best |
| v6.1 | 0.3 | 32 | 4K + SFT×46 | ~6h | Math +0.8 | 已完成 Done |
| v7 (latest) | 0.05 | 32 | 4K + SFT×1500 | ~6h | Code +0.3 ↑ | 最新版本 Latest |
CE (语言建模) + Feature Matching (VAA 对齐) + KL (软标签蒸馏) CE (Language Modeling) + Feature Matching (VAA Alignment) + KL (Soft-Label Distillation)
arXiv: 2602.14301Proof-of-Contribution 机制
Proof-of-Contribution Mechanism
基于贡献度的积分奖励
Points based on contribution quality
获取改进的全局 MoE
Get the improved Global MoE
平台记录的贡献历史
Platform-recorded contribution history
加入 DeepFusion 网络，与全球 AI 一起进化
Join the DeepFusion network and evolve with global AI