JetMoE: Reaching LLaMA2 Performance with 0.1M Dollars
Key Messages
- JetMoE-8B is trained with less than $ 0.1 million1 cost but outperforms LLaMA2-7B from Meta AI, who has multi-billion-dollar training resources. LLM training can be much cheaper than people generally thought.
- JetMoE-8B is very open and academia-friendly because:
- It only uses public datasets for training, and the code is open-sourced. No proprietary resource is needed.
- It can be finetuned with very limited compute budget (e.g., consumer-grade GPU) that most labs can afford.
- JetMoE-8B only has 2.2B active parameters during inference, which drastically lowers the computational cost. Compared to a model with similar inference computation, like Gemma-2B, JetMoE-8B achieves constantly better performance.
- Github: https://github.com/myshell-ai/JetMoE
- HuggingFace: https://huggingface.co/jetmoe/jetmoe-8b
- Chat Demo on Lepton AI: https://www.lepton.ai/playground/chat?model=jetmoe-8b-chat
- Technical report: https://arxiv.org/pdf/2404.07413.pdf
Authors
The project is contributed by Yikang Shen, Zhen Guo, Tianle Cai and Zengyi Qin. For technical inquiries, please contact Yikang Shen. For media and collaboration inquiries, please contact Zengyi Qin.
Collaboration
If you have great ideas but need more resources (GPU, data, funding, etc.), welcome to contact Zengyi Qin. We are open to collaborations and are actively supporting high-quality open-source projects.
Benchmarks
We use the same evaluation methodology as in the Open LLM leaderboard. For MBPP code benchmark, we use the same evaluation methodology as in the LLaMA2 and Deepseek-MoE paper. The results are shown below:
| Model | Active Params | Training Tokens | MBPP | Open LLM Leaderboard Average | ARC | Hellaswag | MMLU | TruthfulQA | WinoGrande | GSM 8K |
|---|---|---|---|---|---|---|---|---|---|---|
| Gemma-2B | 2B | 2T | 28.0 | 46.4 | 48.4 | 71.8 | 41.8 | 33.1 | 66.3 | 16.9 |
| DeepseekMoE-16B | 2.8B | 2T | 34.0 | 51.1 | 53.2 | 79.8 | 46.3 | 36.1 | 73.7 | 17.3 |
| LLaMA2-7B | 7B | 2T | 20.8 | 51.0 | 53.1 | 78.6 | 46.9 | 38.8 | 74.0 | 14.5 |
| LLaMA-13B | 13B | 1T | 22.0 | 51.4 | 56.2 | 80.9 | 47.7 | 39.5 | 76.2 | 7.6 |
| JetMoE-8B | 2.2B | 1.25T | 34.2 | 53.0 | 48.7 | 80.5 | 49.2 | 41.7 | 70.2 | 27.8 |
| Model | MT-Bench Score |
|---|---|
| GPT-4 | 9.014 |
| GPT-3.5-turbo | 7.995 |
| Claude-v1 | 7.923 |
| JetMoE-8B-chat | 6.681 |
| Llama-2-13b-chat | 6.650 |
| Vicuna-13b-v1.3 | 6.413 |
| Wizardlm-13b | 6.353 |
| Llama-2-7b-chat | 6.269 |
To our surprise, despite the lower training cost and computation, JetMoE-8B performs even better than LLaMA2-7B, LLaMA-13B, and DeepseekMoE-16B. Compared to a model with similar training and inference computation, like Gemma-2B, JetMoE-8B achieves better performance.
Model Details
JetMoE uses a sparsely activated architecture inspired by ModuleFormer. JetMoE-8B has 24 blocks. Each block has two MoE layers: Mixture of Attention heads (MoA) and Mixture of MLP Experts (MoE). Each MoA and MoE layer has 8 expert, and 2 experts are activated for each input token. It has 8 billion parameters in total and 2.2B active parameters. JetMoE-8B is trained on 1.25T tokens from publicly available datasets, with a learning rate of 5.0 x 10-4 and a global batch-size of 4M tokens.
Training Details
Our training recipe follows the MiniCPM's two-phases training method. Phase 1 uses a constant learning rate with linear warmup and is trained on 1 trillion tokens from large-scale open-source pretraining datasets, including RefinedWeb, Pile, Github data, etc. Phase 2 uses exponential learning rate decay and is trained on 250 billion tokens from phase 1 datasets and extra high-quality open-source datasets. We used a 96×H100 GPU cluster for 2 weeks to train the model.
Acknowledgement
We express our gratitude to Shengding Hu for his valuable advice on the Phase 2 data mixture. We also express our gratitude to Exabits for their assistance in setting up the GPU clusters, and to Lepton AI for their support in setting up the chat demo.