steer-like-the-llm

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

steer-like-the-llm

★ 9 Open GitHub ↗

(ICML'26 spotlight): Steer Like the LLM: Activation Steering that Mimics Prompting

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Steer Like the LLM

This repository contains the code for our paper Steer Like the LLM: Activation Steering that Mimics Prompting.

## Setup

### Environment variables
First configure the following environment variables:
- `$HF_TOKEN` should contain a Huggingface Hub token.
- `$OPENAI_API_KEY` should contain an OpenAI API key.
- `$ASE_DATA_DIR` (optional) controls where datasets, activation steering checkpoints, experiment results will be stored. By default these are stored under this repo in a `data` subfolder.

### Dependencies

You can install the dependencies with:
```bash
./setup.sh
```
The script assumes that python3 points a python binary for version 3.12 or higher. It will:
- install a library `ase` (activation steering experiments) in a virtual environment (`.venv` by default)
- clone the `persona_vectors` repository (its dependencies are already installed in the virtual environment)
- clone the `instruction_following_eval` and `llm-steer-instruct` repos which are used to create the results on the `IFEval` version used in Stolfo et al. 2025.
- clone the `axbench` repo and install its dependencies in a separate virtual environment.

After installation activate the virtual environment `source .venv/bin/activate`

### Benchmark data
To prepare the necessary data for evaluating on persona vectors for a given LLM run the following command in the virtual environment that was created in the previous step:

```bash
python scripts/prepare_persona_vectors_data.py --model_name meta-llama/Llama-3.2-3B-Instruct --gpu 0
```

To prepare the data for IFEval similar to LLM Instruct paper (Stolfo et al. 2025) use:
```bash
python scripts/prepare_llm_instruct_data.py
```

For Axbench no separate script is needed.

## Repo structure

- `src/ase` contains code that is reused across different experiments. Under `src/ase/steering` you find the modules for steering.
- `experiments` contains the scripts and notebooks for running the paper experiments and analyzing the results. The .sh scripts can be executed with QUICK_TEST=1 (e.g., `QUICK_TEST=1 ./experiments/axbench/run_eval.sh`) for a quick dry run. Scripts with `eval` in the name train steering modules on training data and evaluate them on the evaluation split. Scripts with `tune` in the name do hyperparameter tuning, they evaluate and train on different data.

## Note on terminology

Some naming conventions are different from the paper:
- The PSR architecture is refered to as `Focused`
- The MSE loss is often refered to as `psi` or `Prompt Steering Imitation`.

steer-like-the-llm

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

steer-like-the-llm

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README