|
--- |
|
library_name: transformers |
|
license: llama3.1 |
|
base_model: |
|
- meta-llama/Llama-3.1-8B-Instruct |
|
tags: |
|
- nlrl |
|
--- |
|
|
|
# Model Card for Llama-3.1-8B-Instruct-NLRL-TicTacToe-Policy |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
- **Developed by:** NLRL Team |
|
- **Model type:** Language Policy Model for TicTacToe |
|
- **Language(s):** English |
|
- **License:** MIT |
|
- **Finetuned from model:** LLaMA-3.1-8B-Instruct |
|
|
|
This model serves as a language policy in Natural Language Reinforcement Learning (NLRL) framework, specifically trained for the TicTacToe game. It generates actions through chain-of-thought reasoning and outputs move decisions. |
|
|
|
## Uses |
|
|
|
### Direct Use |
|
This model can be used as a TicTacToe player that explains its strategic thinking through natural language before making moves. The model generates both reasoning chains and final move decisions. |
|
|
|
### Out-of-Scope Use |
|
This model is specifically trained for TicTacToe and should not be used for other games or tasks. |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
Training data consists of state-action pairs collected through NLRL actor-critic learning process, with language-based Monte Carlo value estimates used for policy improvement. |
|
|
|
### Training Procedure |
|
- Trained using FSDP (Fully Sharded Data Parallel) across 4 H100 GPUs |
|
- Learning rate: 1e-5 |
|
- Training epochs per iteration: 2 |
|
- Batch size: 8 |
|
- Max sequence length: 1024 |
|
|
|
## Evaluation |
|
- Tested against deterministic (first-move) and random opponent strategies |
|
- Achieves >90% win rate against both opponent types after convergence |
|
|
|
## Model Architecture |
|
- Base model: LLaMA-3.1-8B-Instruct |
|
- Input: Text description of TicTacToe board state |
|
- Output: Chain-of-thought reasoning followed by move decision |
|
|
|
## Citation |
|
```bibtex |
|
@misc{nlrl, |
|
title={Natural Language Reinforcement Learning}, |
|
author={Xidong Feng and Ziyu Wan and Haotian Fu and Bo Liu and Mengyue Yang and Girish A. Koushik and Zhiyuan Hu and Ying Wen and Jun Wang}, |
|
year={2024}, |
|
eprint={2411.14251}, |
|
archivePrefix={arXiv}, |
|
primaryClass={cs.LG}, |
|
url={https://arxiv.org/abs/2411.14251}, |
|
} |
|
``` |
|
|
|
## Model Card Contact |
|
benjaminliu.eecs@gmail.com |
