Setup

Install uv. Then run:

uv sync

This installs all dependencies.

Download Monte Carlo Estimates

hf download rangell/mc_estimates --repo-type dataset --local-dir monte_carlo_estimates/results/

Monte Carlo estimation

StrongREJECT

uv run --module monte_carlo_estimates.src.strong_reject.generate_mc_est --target-model meta-llama/Llama-3.1-8B-Instruct --num-return-sequences 10000 --output-dir monte_carlo_estimates/results/strong_reject/

Persona Vectors

uv run --module persona_vectors.eval.eval_persona --model meta-llama/Llama-3.1-8B-Instruct --trait sycophantic --output_path monte_carlo_estimates/results/persona_vectors/sycophantic/Llama-3.1-8B-Instruct_mc_est_10k.csv --version eval --n_per_question 10000 --overwrite True
uv run --module monte_carlo_estimates/src/persona_vectors/reformat_output.py --trait sycophantic --csv_infile monte_carlo_estimates/results/persona_vectors/sycophantic/Llama-3.1-8B-Instruct_mc_est_10k.csv monte_carlo_estimates/results/persona_vectors/sycophantic/Llama-3.1-8B-Instruct_mc_est_10k.json

Steering Vector Computation

Refusal Direction

Even if you only want to use persona vectors, you need to compute the refusal direction

source .venv/bin/activate
cd refusal_direction
python -m pipeline.run_pipeline --model meta-llama/Llama-3.1-8B-Instruct

Persona Vectors

To generate the persona vectors, run the following

uv run --module persona_vectors.gen_vec_pipeline --model meta-llama/Llama-3.1-8B-Instruct --trait sycophantic

The above pipeline also does a little search over steering coefficients and layers to find which values work the best.

Rare Event Estimation

Estimating harm example

uv run --module smc.est_unsafe_to_unsafe --model_name meta-llama/Llama-3.1-8B-Instruct  --num_particles 100 --fwd_batch_size 100 --max_new_tokens 150 --use_cem

Estimating persona example

uv run --module smc.est_persona --model_name meta-llama/Llama-3.1-8B-Instruct --num_particles 100 --fwd_batch_size=100 --max_new_tokens 1000 --trait sycophantic --steering_type response --use_cem