cfierro/llama-3.1-8b-fft-othello-multi-format-1e-5-2m
The cfierro/llama-3.1-8b-fft-othello-multi-format-1e-5-2m is an 8 billion parameter Llama-3.1-8B-Instruct model fine-tuned by cfierro. This model is specifically trained on 2 million Othello game data points in a multi-format setup, without chat replay, to analyze data scaling for valid move generation. It focuses on understanding how many games an 8B model needs to achieve near-perfect valid move rates in Othello. The model is intended for research into game-playing LLMs and data efficiency, particularly for Othello game mechanics.
Loading preview...
Model Overview
This model, cfierro/llama-3.1-8b-fft-othello-multi-format-1e-5-2m, is an 8 billion parameter Llama-3.1-8B-Instruct variant, fine-tuned by cfierro. It was trained using Axolotl on a specialized dataset of 2 million Othello games, presented in a multi-format style. A key differentiator is the exclusion of chat replay data, focusing solely on game mechanics to isolate the signal for valid move generation.
Key Capabilities & Purpose
- Othello Game Play Analysis: Designed to investigate the data scaling requirements for an 8B model to achieve high accuracy in generating valid Othello moves.
- Data Efficiency Probe: Serves as a probe to determine the optimal number of game examples needed for robust Othello play, comparing against previous runs that included chat replay.
- Multi-Format Training: Utilizes a multi-format representation of Othello games, enhancing its ability to process diverse game states.
Training Details
The model underwent full fine-tuning for 22,000 steps, processing approximately 2 million unique Othello games. The training process generated a "checkpoint ladder," allowing for analysis of performance at various data scales (e.g., 176k, 500k, 1M, 2M games). This setup enables researchers to observe the valid-move rate curve as data increases. The training explicitly masked the initial game prefix, focusing the loss calculation on the formatted moves and the end-of-text token. The model's performance is primarily evaluated by its valid-move rate, as the absence of chat replay means it is not intended for conversational or verbalization tasks.