steering-eval

Evaluating steering vectors across diverse tasks.

This repository contains the code and data from Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks. In particular, this code supports evaluations of Function Vectors (Todd et al.) and In-Context Vectors (Liu et al.).

Setup

We suggest creating a conda environment to manage the required packages:

conda create --name steer-eval python
conda activate steer-eval
pip install -r requirements.txt

Data

All data used in our experiments can be found in /function_vectors/dataset_files/abstractive. We broadly categorize datasets into two main categories: functional and behavioral. See our paper for more information on tasks we test.

Evaluation

The scripts used to evaluate steering vectors are functional_experiment.py for vanilla functional tasks (zero-shot and shuffle-label few-shot), natural_text_experiment.py for out-of-distribution natural text functional tasks, behavioral_experiment.py for behavioral tasks, and ablation_experiment.py for our ablation experiments. All take the following arguments:

• --model_name: Huggingface model name. Supports Llama and Pythia models.
  • --hf_token: Optional, huggingface token for gated models
• --dataset_name: Dataset file name
• --run_clean: Set this flag to evaluate and save clean model performance
• --run_icv: Set this flag to evaluate and save ICV-steered performance
• --run_fv: Set this flag to evaluate and save FV-steered performance
• --save_path_root: Directory to store results
• --device: Which device to use (cpu or cuda)
• --save_cie: Set this flag to save causal indirect effect calculated for FV extraction
• --demos: List of demonstration dataset sizes to use for extracting in-context vectors
  • Usage: --demos 5 50 100 150
• --alphas: List of in-context vector strengths to evaluate
  • Usage: --alphas 1.0 1.5
• --seeds: Number of times to run experiment with random seed
• --specific_seeds: List of specific seeds to run

You can pass in custom experiment parameters for each type of evaluation script as follows:

• natural_text_experiment.py:
  • --n_eval_templates: Number of natural text prompt formats to test. Code samples templates randomly without replacement. Max 5 (3 for country-capital task)
• functional_experiment.py:
  • --shots: List of shots to evaluate over
    • Usage: --shots 0 3
• ablation_experiment.py:
  • --find_layers: Keyword for layers to edit. Options are:
    • "first": first layer of the model
    • "last": last layer
    • "all": all layers
    • "middle": middle layer of model
    • "middle2": middle two layers of model
    • "middle4": middle four layers of model