This evaluates LLMs trained on PDN format Checkers games through the use of linear probes. We can check the LLMs internal understanding of board state.
This repo can train, evaluate, and visualize linear probes on LLMs that have been trained to play Checkers with PDN strings.
python model_setup.py
Split the dataset for training the probes
python data/split_traintest.py
The train_test_checkers.py script can be used to either train new linear probes or test a saved probe on the test set.
Command line arguments:
--mode: Specifies train or test. Optional, defaults to train.
--probe: Determines the type of probe to be used. piece probes for the piece type on each square. Optional, defaults to piece. We only have piece in Checkers
Examples:
Train piece board state probes:
python train_test_checkers.py
See all options: python train_test_checkers.py -h
Run all cells in the chess_llm_interpretability/probe_output_visualization.ipynb for visualizing the probe_outputs.
Much of my linear probing was developed using Neel Nanda's linear probing code as a reference. Here are the main references used:
https://colab.research.google.com/github/neelnanda-io/TransformerLens/blob/main/demos/Othello_GPT.ipynb https://colab.research.google.com/github/likenneth/othello_world/blob/master/Othello_GPT_Circuits.ipynb https://www.neelnanda.io/mechanistic-interpretability/othello https://github.com/likenneth/othello_world/tree/master/mechanistic_interpretability