Transparent LLM alignment for all.
Hi, I am Magpie 🐦, your efficient and high-quality synthetic data generation pipeline!
🕸️ Project Website: https://magpie-align.github.io/
📄 Arxiv: https://arxiv.org/abs/2406.08464 (Accepted by ICLR 2025)
🤗 HF Paper Page: https://huggingface.co/papers/2406.08464
😬 Codes: https://github.com/magpie-align/magpie
🤗 Magpie Demo: https://huggingface.co/spaces/davanstrien/magpie (Thanks a lot for the implementation from @davanstrien!)
🐦 MagpieLM: MagpieLM-4B, MagpieLM-8B
Questions? Please contact Zhangchen and/or Yuchen by email or raise an issue in Github.
Here are some filtered datasets made by the authors, which are utilized in our Magpie-Align models. We also encourage you to create and apply your own filters to customize datasets.
We've kept these datasets within the 200K-300K range for your convenience. We found this range represents a sweet spot balancing model performance and training time.
The full list of filtered datasets can be found here.
| Model Name | Dataset | Size | Type | Description |
|---|---|---|---|---|
| Llama 3.1 70B Instruct | Magpie-Llama-3.1-Pro-MT-300K-Filtered | 300K | SFT | (🌟 Flexible License! 🌟) Select 300K high quality multi-turn conversations from Magpie-Llama-3.1-Pro-MT-500K. |
| Llama 3 70B Instruct | Magpie-Pro-300K-Filtered | 300K | SFT | Apply a filter and select 300K high quality conversations from Magpie-Pro-1M. |
| Llama 3 70B Instruct | Magpie-Pro-MT-300K | 300K | SFT | Select 300K difficult questions from Magpie-Pro-1M and extend to multi-turn conversations. |
| Llama 3 70B Instruct | Magpie-Reasoning-150K | 150K | SFT | Reasoning booster with 150K math + code + reasoning conversations. Recommend mixing with Magpie-Pro-MT-300K. |
| Qwen2 72B Instruct | Magpie-Qwen2-Pro-200K-Chinese | 200K | SFT | Apply a filter and select 200K high quality Chinese conversations from Magpie-Qwen2-Pro-1M. |
| Gemma-2-27b-it | Magpie-Gemma2-Pro-200K-Filtered | 200K | SFT | (🌟 Flexible License! 🌟) Apply a filter and select 200K conversations from Magpie-Gemma2-Pro-534K. |
| Llama 3 8B Instruct | Magpie-Air-DPO-100K | 100K | DPO | DPO dataset via Best-of-N sampling and rewards. |
