This repository contains the code for our paper, Representing Rule-based Chatbots with Transformers. The code can be used to generate synthetic ELIZA training data, train and evaluate Transformers on ELIZA transcripts, and conduct some analysis of the learned mechanisms. Please see our paper for more details.
Install PyTorch and then install the remaining requirements: pip install -r requirements.txt.
This code was tested using Python 3.12 and PyTorch version 2.2.2.
The code we used to generate the data is in src/generate_data.py. The scripts directory contains the configurations for the datasets we used in our paper.
- Multi-turn conversations: scripts/generate_multi_turn_data.sh
- Single-turn conversations, varying the amount of repetition in the copying segments: scripts/generate_single_turn_data.sh
The datasets we used in our experiments can also be downloaded directly from HuggingFace via this link: https://huggingface.co/datasets/danf0/eliza.
To train Transformers on ELIZA conversations, see src/run.py. The scripts directory contains the configurations for the datasets we used in our paper, for multi-turn conversations and single-turn conversations. We ran our experiments on NVIDIA GeForce RTX 2080 Ti GPUs with 11GB of memory. The single-turn models were trained for approximately 12 hours each, and the multi-turn models were trained for approximately 48 hours.
The training code saves predictions on the validation set throughout training. src/utils/analysis_utils.py contains a number of utilities for analyzing these predictions. These can be used to measure performance as a function of various properties of the prediction, such as the turn type, the template, and the state of the memory queue.
src/probing.py contains code examinging attention embeddings, to investigate whether attentions scores are influenced more by content or position. To run this analysis with the configuration used in the paper, see scripts/run_attention_probe.sh.
We generated counterfactual datasets to examine the memory mechanisms learned by the models we trained.
The code for generating these datasets is in src/counterfactuals.py.
The datasets we used in our experiments can also be downloaded directly from HuggingFace (see https://huggingface.co/datasets/danf0/eliza).
After generating a counterfactual dataset, a model can be evaluated using src/run.py, using the --eval_only flag.
If you have any questions about the code or paper, please email Dan (dfriedman@cs.princeton.edu) or open an issue.
@article{friedman2024representing,
title={Representing Rule-based Chatbots with Transformers},
author={Friedman, Dan and Panigrahi, Abhishek and Chen, Danqi},
year={2024}
}