Differential Preference Steering (DPS)

Official implementation of Differential Preference Steering (DPS). DPS detects preference-sensitive attention heads and steers them during decoding to personalize LLM outputs.

Setup

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Data

Datasets are not included. Point LAMP_DATA_ROOT at a directory with one subdirectory per task (LaMP-1/ … LongLaMP-4/), or run python scripts/prefetch_lamp_datasets.py to download the raw files. LaMP loading, prompting, retrieval, and metrics are implemented in src/lamp_benchmark/.

Usage

Run a Hydra-configured evaluation:

python scripts/main.py data=lamp_1 model=llama3_8b_instruct decoder=dps

Build preference-head artifacts:

python preference_head/cluster_profiles.py --task LaMP-1 --split dev --k 25 \
  --output_dir artifacts/cluster_runs/lamp1_k25 --save_embeddings
python preference_head/preference_head_detection.py --task LaMP-1 --split dev \
  --save_dir artifacts/preference_heads

Run weighted cluster-routed DPS:

python scripts/run_weighted_dps.py --task LaMP-1 \
  --cluster_file artifacts/cluster_runs/lamp1_k25/clusters.json \
  --cluster_heads_dir artifacts/cluster_heads/lamp1_k25 \
  --embeddings_file artifacts/cluster_runs/lamp1_k25/embeddings.npy

Layout

• src/ — datasets, metrics, model wrappers, and decoding methods (src/lamp_benchmark/ holds the LaMP utilities).
• preference_head/ — profile clustering and preference-head detection.
• configs/ — Hydra configs for data, models, and decoders.
• scripts/ — entry points for evaluation, weighted DPS, and data prefetch.