In-Context Learning Dynamics with Random Binary Sequences

This repo contains code + data for the ICLR 2024 paper In-Context Learning Dynamics with Random Binary Sequences

Files

• Analysis.ipynb - Main analysis notebook generating all plots in the paper for OpenAI models.

• Analysis-HuggingFace.ipynb - Generate plots for open-source huggingface LLMs.

• ModelSelection.ipynb - Minimal example of Bayesian model selection described in Figure 2 and Appendix D.

• DistractMMLU.ipynb - MMLU "distraction" task used for Appendix E.

• MinimalExample.ipynb - Minimal example of going from prompt, to querying openAI, to generating plots with Randomness Judgment task.

• DataPreProcessing.ipynb - Records how I converted raw openAI query data to the pickle objects loaded in Analysis.ipynb. I didn't add imports and haven't tested this like the other notebooks.

• utils.py - Utilities for plotting and processing data, mostly used in Analysis.ipynb.

• hf.py - Code for running inference on HuggingFace transformers models, and extracting token logits.

• collect_data.py - Script for querying OpenAI with all data for randomness generation. Set arg type to 'flips_random' for generation with varying $p(Tails)$, or to tree_formal for generation with formal concept learning.

• collect_hf.py - Script for running inference on huggingface models and collecting outputs.

Data

Data is available for download at TODO.

Requirements

See requirements.txt for python package requirements.

Querying OpenAI requires the batch_prompt package. The plotting code depends on an earlier version of this, which can be installed by:

git clone [email protected]:ebigelow/batch-prompt.git    # clone repo
git checkout 15925fb                                  # checkout specific earlier version
ln -s ./batch_prompt ../batch_prompt                  # add a symbolic link so package can be imported to notebooks, etc.

Citation

Bigelow, E. J., Lubana, E. S., Dick, R. P., Tanaka, H., & Ullman, T. D. (2023). In-Context Learning Dynamics with Random Binary Sequences. In the proceedings of the 12th International Conference on Learning Representations (ICLR).

Contact

If you encounter problems, submit an issue on github or contact me (see paper for email).