🔍 Induction-Gram 🔍

Interpretable Language Modeling via Induction-head Ngram Models (Kim*, Mantena* et al. 2024).

Recent large language models (LLMs) have excelled across a wide range of tasks, but their use in high-stakes and compute-limited settings has intensified the demand for interpretability and efficiency. We address this need by proposing Induction-head ngram models (Induction-Gram), a method that builds an efficient, interpretable LM by bolstering modern ngram models with a hand-engineered "induction head". This induction head uses a custom neural similarity metric to efficiently search the model's input context for potential next-word completions. This process enables Induction-Gram to provide ngram-level grounding for each generated token. Moreover, experiments show that this simple method significantly improves next-word prediction over baseline interpretable models (up to 26%p) and can be used to speed up LLM inference for large models through speculative decoding. We further study Induction-Gram in a natural-language neuroscience setting, where the goal is to predict the next fMRI response in a sequence. It again provides a significant improvement over interpretable models (20% relative increase in the correlation of predicted fMRI responses), potentially enabling deeper scientific investigation of language selectivity in the brain.

Setup

• Clone the repo and run pip install -e . to install the alm package locally
• Set paths/env variables in alm/config.py to point to the correct directories to store data
• Experiments
  • For language modeling experiments, please refer to experiments/readme.md.

Organization

• alm is the main package
  • alm/config.py is the configuration file that points to where things should be stored
  • alm/data contains data utilities
  • alm/models contains source for models
• data contains code for preprocessing data
• experiments contains code for language modeling experiments

Citation

If you find this work useful, please cite:

@misc{kim2024inductiongram,
      title={Interpretable Language Modeling via Induction-head Ngram Models}, 
      author={Eunji Kim and Sriya Mantena and Weiwei Yang and Chandan Singh and Sungroh Yoon and Jianfeng Gao},
      year={2024},
      eprint={2411.00066},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2411.00066}, 
}