ReGPT

Probing Reasoning of Language Models with Inductive In-Context Learning

ReGPT is used to evaluate the hypothesis of learning to associate Facts (in this case are examples that match a regex) to Rules (in this case it is a Regex). The advantage of ReGPT over previous evaluation methods is that there is a ground-truth on whether a Fact can be described by a Rule (for example we can evaluate membership via re.match).

Description

Illustration on the differences between our approach in Training (left) and Evaluating (right) as compared to previous methods. For ReGPT we inject a Rule (R) during training (bottom row of the red outline) as compared to GPT that is trained with additional Facts (F). Previous approaches directly optimize for a Rule synthesis task (biased) or use tests that can be inapplicable for LMs (pink outline). During evaluation, we probe ReGPT with Inductive In-Context Learning (blue outline) to infer a Rule ($\hat{R}$) from Facts that we use to generate additional Facts (blue tokens). ReAnalogy evaluate whether the generated Facts abide by the ground-truth Rule. Our benchmark can implicitly evaluate on whether a model has learned to infer the underlying Rules as opposed to explicitly evaluate a model optimized for Rule Synthesis. For Inductive In-Context Learning the soundness of the Rule is unambiguous and as such it improves performance when compared to a baseline model.

Usage

Install

git clone https://github.com/fostiropoulos/regpt
pip install -e regpt

Train

python -m regpt  \
[--dataset {reanalogy,deep,kb13}] \
--dataset_path DATASET_PATH \ # location of ReAnalogy
--save_path SAVE_PATH \
[--regpt] \ # ReGPT vs GPT
[--n_examples {2,5,8,12}] \
[--n_layers {2,6,24}] \
[--batch_size BATCH_SIZE] \
[--device {cuda:0,cuda:1,cpu}] \
[--resume]

Evaluate

from regpt.compiler import Compiler
# chkpt_path must be ReGPT_***.pt saved in the SAVE_PATH
c = Compiler(dataset_path, chkpt_path)
c.evaluate(n_samples=1)

ReAnalogy

ReAnalogy is a dataset composed of regular expressions that is used to train ReGPT.

Cite

@inproceedings{fostiropoulos2023probing,
  title={Probing Reasoning of Language Models with Inductive In-Context Learning},
  author={Fostiropoulos, Iordanis and Itti, Laurent},
  booktitle={International Joint Conference on Artificial Intelligence 2023 Workshop on Knowledge-Based Compositional Generalization},
  year={2023}
}

fostiropoulos / ReGPT