This is the official repository for "Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing". Magpie generates high-quality alignment data by prompting aligned LLMs with their pre-query templates. Unlike many existing synthetic data generation methods, Magpie doesn't rely on prompt engineering or seed questions for generating synthetic data. Instead, it uses the prompt template of an aligned LLM to generate both the user query and an LLM response.
- π€ Huggingface (Models and Datasets)
- π§ Dataset Navigation
- πΈοΈ Website
- π Technical Report
- π€ Magpie Demo (Thanks a lot for the implementation from @davanstrien!)
- π¦ Chat with Magpie
- [2024/10/20] Magpie Qwen2.5 dataset is out! 1M from Qwen2.5 72B!
- [2024/09/17] Ship two new models with SOTA performance: πΌππππππ»πΌ-π²πππ (4B & 8B)! See collection here!
- [2024/08/19] Three preference optimization datasets, Magpie-Air-DPO-100K-v0.1, Magpie-Pro-DPO-100K-v0.1, and Magpie-Llama-3.1-Pro-DPO-100K-v0.1 are out!
- [2024/07/25] Magpie Llama-3.1 dataset is out! 1M from Meta-Llama-3.1-70B-Instruct! More friendly license compared with Llama-3 π!
- [2024/07/21] Magpie Gemma2 dataset is out! 534K from Gemma-2-27b-it!
- [2024/07/19] Llama-3-8B-Magpie-Align-v0.3 is out with enhanced Chinese question-answering ability, thanks to our new Chinese instruction dataset!
- [2024/07/14] Llama-3-8B-Magpie-Align-v0.2 is out with enhanced reasoning ability, thanks to our new reasoning booster dataset!
- [2024/07/04] Magpie Qwen2 dataset is out! 1M from Qwen2 72B and 3M from Qwen2 7B.
- [2024/07/03] π Our open aligned model, Llama-3-8B-Magpie-Align-v0.1 is out! It is π the best <30B Model in AI2 WildBench Leaderboard! Even better than the official Meta-Llama-3-8B-Instruct model!
- [2024/06/24] Magpie Phi 3 dataset is out! 1M from Phi 3 Medium.
- [2024/06/12] Magpie Llama-3 dataset is out! 1M from Llama-3 70B and 3M from Llama-3 8B.
- [2024/06/12] Magpie technical report is out! Let's make high-quality alignment data open for all!
Currently, Magpie has been tested on the Llama-3, Qwen2, Phi 3 and Gemma-2 series. Please submit an issue for more model support.
| Model Family | Magpie | Magpie Scripts | Datasets | Size |
|---|---|---|---|---|
| Llama 3.1 | β * | 8B,70B | 70B,405B(Argilla) | 1M |
| Llama 3 | β | 8B,70B | 8B,70B | 3M + 1M |
| Qwen2.5 | β | 3B,7B,14B,32B,72B | 72B | 1M |
| Qwen2 | β | 7B,72B,Math 7B | 7B,72B | 3M + 1M |
| Phi 3 | β | mini,small,medium | medium | 1M |
| Gemma-2 | β ** | 9B,27B | 27B | 534K |
| Gemma-1.1 | βοΈ | 7B | ||
| Llama 2 | βοΈ | 7B,70B | ||
| Vicuna | βοΈ | 7B | ||
| Mistral | βοΈ | 7B | ||
| Yi | βοΈ | 34B | ||
| DeepSeek Coder | βοΈ | Coder V2 Lite |
- β : It works great! (* Apply a logits processor to surpress markdown; ** Apply a filter before generating responses.)
- βοΈ: It works! We can get something interesting, but we may need to design an additional logit processor and/or a filter.
- β: Not work.
- β: Untested.
The navigation of all available Magpie datasets can be found here.
We hope Magpie can contribute to the democratization of AI with enhanced transparency of model alignment processes!
Click Here
High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets. Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis method for generating large-scale alignment data named Magpie. Our key observation is that aligned LLMs like Llama-3-Instruct can generate a user query when we input only the left-side templates up to the position reserved for user messages, thanks to their auto-regressive nature. We use this method to prompt Llama-3-Instruct and generate 4 million instructions along with their corresponding responses. We perform a comprehensive analysis of the extracted data and select 300K high-quality instances. To compare Magpie data with other public instruction datasets, we fine-tune Llama-3-8B-Base with each dataset and evaluate the performance of the fine-tuned models. Our results indicate that in some tasks, models fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through supervised fine-tuning (SFT) and subsequent feedback learning. We also show that using Magpie solely for SFT can surpass the performance of previous public datasets utilized for both SFT and preference optimization, such as direct preference optimization with UltraFeedback. This advantage is evident on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench.
Build environment
git clone https://github.com/magpie-align/magpie.git
cd magpie
conda create -n magpie python=3.10 -y
conda activate magpie
pip install -r requirements.txt
Get access to Llama-3 models from π€ Huggingface
You can apply for Llama-3 model access here. To login in the terminal, enter:
huggingface-cli login
then enter your Huggingface private key beginning with "hf_".
Play with Jupyter Notebook
The toy example can be found in demo.ipynb. Have fun!
We use Llama-3-8B-Instruct as an example to demonstrate the batched SFT data generation process. To run batched generation, you can simply run:
cd scripts
bash magpie.sh
The script will generate both instructions and responses in the data folder. It has been tested on an RTX 4090 24G GPU. If you are using GPUs with less memory, consider implementing quantization.
We also provide scripts for other models in the scripts folder. You can use this navigation to find specific Magpie scripts. Note that for model sizes greater than 8B, you may need 4*A100 GPUs to run the scripts.
After generating instruction-response pairs, you can extend them to multi-turn conversations. To do so, simply run the following command:
bash magpie-multi-turn.sh ***_ins_res.json
where ***_ins_res.json is the single-turn instruction-response pairs generated in the previous step.
To tag the generated instruction-response pairs, you can run:
cd scripts
bash unitag.sh ***_ins_res.json all
This script will automatically generate quality, difficulty, task category, safety, reward, and language for the generated dataset. You can also generate one tag at a time. For example, if you just want to generate the safety label using device 0, you can run:
cd scripts
bash unitag.sh ***_ins_res.json safety 0
You may generate datasets with different generation configurations. We provide a Jupyter notebook here for concatenating all datasets and converting them to ShareGPT format, which is fully supported by Axolotl for fine-tuning.
Once you have a full dataset converted to ShareGPT format, you can calculate the minimum neighbor distance of each instruction and remove repetitions. To do so, run:
cd exp
python gen_dis.py --input_file ***_sharegpt.jsonl
where ***_sharegpt.jsonl is the dataset path obtained in the previous step. The Python script will take care of building the FAISS index and calculating the minimum distance.
We provide a Jupyter notebook here for simple filtering. You can adjust the filtering parameters to design and apply your own filter based on your needs.
To generate preference data, first prepare filtered instructions following the steps outlined above. For the expected format, please refer to our example here.
Next, please use our provided scripts here to generate multiple responses and compute their corresponding rewards. Finally, your can process the data and upload it to Huggingface using this Jupyter notebook.
Please take a look at the recipes directory for instructions and our Magpie model recipes.
If you find the model, data, or code useful, please cite our paper π€©:
@article{xu2024magpie,
title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
journal={ArXiv},
year={2024},
volume={abs/2406.08464},
url={https://api.semanticscholar.org/CorpusID:270391432}
}