Llama 3 Instruct 8B SPPO Iter3

🚀 Llama-3-Instruct-8B-SPPO-Iter3

This model, Llama-3-Instruct-8B-SPPO-Iter3, is designed for text generation. It uses the Self-Play Preference Optimization method to fine - tune the model, aiming to improve the alignment and performance of the language model.

🚀 Quick Start

This README provides detailed information about the Llama-3-Instruct-8B-SPPO-Iter3 model, including its development, evaluation results, and training hyperparameters.

✨ Features

• Optimized Architecture: Developed based on the meta-llama/Meta-Llama-3-8B-Instruct architecture, with further optimization through Self - Play Preference Optimization.
• Diverse Datasets: Utilizes prompt sets from the openbmb/UltraFeedback dataset, which is split for multiple iterations.
• Synthetic Responses: All responses used during development are synthetic.

📚 Documentation

Model Development

This model was developed using Self-Play Preference Optimization at iteration 3. Starting from the meta-llama/Meta-Llama-3-8B-Instruct architecture, we used the prompt sets from the openbmb/UltraFeedback dataset. The dataset was split into 3 parts for 3 iterations by snorkelai/Snorkel-Mistral-PairRM-DPO-Dataset. All responses used are synthetic.

Links to Other Models

• Llama-3-Instruct-8B-SPPO-Iter1
• Llama-3-Instruct-8B-SPPO-Iter2
• Llama-3-Instruct-8B-SPPO-Iter3

Model Description

Evaluation Results

AlpacaEval Leaderboard Evaluation Results

Open LLM Leaderboard Evaluation Results

Results are reported by using lm-evaluation-harness v0.4.1

Open LLM Leaderboard 2 Evaluation Results

Detailed results can be found here

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e - 07
• eta: 1000
• per_device_train_batch_size: 8
• gradient_accumulation_steps: 1
• seed: 42
• distributed_type: deepspeed_zero3
• num_devices: 8
• optimizer: RMSProp
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_train_epochs: 6.0 (stop at epoch = 1.0)

📄 License

This model is licensed under the Apache - 2.0 license.

🔧 Technical Details

This model is based on the Self - Play Preference Optimization method, which is described in the paper Self-Play Preference Optimization for Language Model Alignment. The method aims to improve the alignment of language models by optimizing the preference of self - play responses.

📚 Citation

@misc{wu2024self,
      title={Self-Play Preference Optimization for Language Model Alignment}, 
      author={Wu, Yue and Sun, Zhiqing and Yuan, Huizhuo and Ji, Kaixuan and Yang, Yiming and Gu, Quanquan},
      year={2024},
      eprint={2405.00675},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}