$ show ericwtodd/function_vectors
This repository implements and evaluates Function Vectors (FVs) for steering large language models (LLMs) without fine-tuning. The core workflow involves computing task-conditioned mean head activations from in-context learning examples, identifying causally influential attention heads via indirect effect, summing their outputs to form the FV, and applying it during inference to elicit target behaviors (e.g., translation, inflection). The system provides a modular Python library with utilities for model loading, prompt construction, activation extraction, intervention, and evaluation, accompanied by a demo notebook and evaluation scripts for running experiments on various linguistic datasets. While the codebase demonstrates clean separation of concerns and contains reusable components for model intervention and prompt handling, it shows signs of technical drift due to dependency rigidity and architecture-specific hooks, and appears to have stagnated in maintenance with no recent updates.
The LLM identified 2 files worth highlighting for reuse:
Provides isolated functions to load and prepare various transformer models (GPT-J, Llama, etc.) for activation intervention. Handles model-specific nuances like setting output_hidden_states and identifying attention layer names in a centralized way. Dependencies are limited to torch and transformers, making it easily extractable for any project requiring LLM intervention.
Implements robust in-context learning prompt construction with special handling for multi-token words via label extension. The module is self-contained (only uses standard libraries and torch) and solves a common problem in ICL where labels may span multiple tokens. Its general approach to prompt templating and token-level labeling makes it reusable across different ICL experiments without modification.