Skip to content

README.md

Latest commit

 

History

History
197 lines (169 loc) · 6.57 KB

README.md

File metadata and controls

197 lines (169 loc) · 6.57 KB

User Modeling

Hugging Face Collection

This project contains code to generate the core data for studying user modeling.

See Scalably Extracting Latent Representations of Users for more details.

Note that the datasets for Llama3.1-8B-Instruct and Llama3.1-70B-Instruct are already on HuggingFace (see the collection). The commands below can reproduce them, but are primarily for if you want to run belief evaluation for new models.

Setup

Follow the instructions in the Installation section of the README in the root of the repo. Then install and activate the user_modeling environment by running

luce install user_modeling
luce activate user_modeling

This will cd to the user_modeling project folder project/user_modeling, which is where all scripts should be executed from.

Define Constants

Replace HF_MODEL_ID and HF_MODEL_NAME with the appropriate values.

export DATA_DIR=<...>
export PORT=2733
export HF_MODEL_ID=meta-llama/Llama-3.1-8B-Instruct
export HF_MODEL_NAME=llama31_8b
export OPENAI_API_KEY=<...>
export ANTHROPIC_API_KEY=<...>
export HF_TOKEN=<...>

SynthSys

Generate System and User Prompts

Generate user prompts using o3:

python scripts/generate_user_prompts.py \
    --db_dir $DATA_DIR/user_prompts_db \
    --min_num_prompts 100 \
    --num_candidates_per_call 100 \
    --generation_provider openai \
    --generation_model_name o3 \
    --generation_max_new_tokens 10_000

Generate user prompts using Claude Opus 4:

python scripts/generate_user_prompts.py \
    --db_dir $DATA_DIR/user_prompts_db \
    --min_num_prompts 100 \
    --num_candidates_per_call 100 \
    --generation_provider anthropic \
    --generation_model_name claude-opus-4-1-20250805 \
    --generation_max_new_tokens 5_000

Generate system prompts using o3:

python scripts/generate_system_prompts.py \
    --db_dir $DATA_DIR/system_prompts_db \
    --min_num_prompts 100 \
    --num_candidates_per_call 100 \
    --provider openai \
    --model_name o3 \
    --max_new_tokens 10_000

Generate system prompts using Claude Opus 4:

python scripts/generate_system_prompts.py \
    --db_dir $DATA_DIR/system_prompts_db \
    --min_num_prompts 100 \
    --num_candidates_per_call 100 \
    --provider anthropic \
    --model_name claude-opus-4-1-20250805 \
    --max_new_tokens 5_000

1. Launch Faithfulness (Belief) Evaluation

Start vLLM server for generating from the subject model:

vllm serve $HF_MODEL_ID \
    --tensor-parallel-size <number of gpus> \
    --host 0.0.0.0 \
    --port $PORT

For the train split:

python scripts/run_synthsys_faithfulness_eval.py \
    --hf_model_id $HF_MODEL_ID \
    --num_subject_completions 2 \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/train_faith_db \
    --system_prompt_db_dir $DATA_DIR/system_prompts_db \
    --user_prompt_db_dir $DATA_DIR/user_prompts_db \
    --split train \
    --subject_gen_client_url http://0.0.0.0:$PORT/v1 \
    --num_examples 1_000_000

For the test split:

python scripts/run_synthsys_faithfulness_eval.py \
    --hf_model_id $HF_MODEL_ID \
    --num_subject_completions 2 \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/test_faith_db \
    --system_prompt_db_dir $DATA_DIR/system_prompts_db \
    --user_prompt_db_dir $DATA_DIR/user_prompts_db \
    --user_prompt_type user_prompt \
    --split test \
    --subject_gen_client_url http://0.0.0.0:$PORT/v1 \
    --num_examples 10_000

2. Evaluate Baseline Beliefs (No System Prompt)

Run this if you want to evaluate all available user prompts in $DATA_DIR/user_prompts_db:

python scripts/run_synthsys_faithfulness_eval.py \
    --hf_model_id $HF_MODEL_ID \
    --num_subject_completions 2 \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/baseline_faith_db \
    --user_prompt_db_dir $DATA_DIR/user_prompts_db \
    --user_prompt_type user_prompt \
    --subject_gen_client_url http://0.0.0.0:$PORT/v1 \
    --eval_baseline True

If you only want to evaluate user prompts from train/test faithfulness eval:

python scripts/run_synthsys_faithfulness_eval.py \
    --hf_model_id $HF_MODEL_ID \
    --num_subject_completions 2 \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/baseline_faith_db \
    --user_prompt_db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/<train_faith_db or test_faith_db> \
    --user_prompt_type faith_result \
    --subject_gen_client_url http://0.0.0.0:$PORT/v1 \
    --eval_baseline True

3. Aggregate Faithfulness Eval Results

For the train split:

python scripts/run_subject_input_aggregation.py \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/train_subject_input_db \
    --faithfulness_eval_db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/train_faith_db \
    --baseline_faithfulness_eval_db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/baseline_faith_db

For the test split:

python scripts/run_subject_input_aggregation.py \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/test_subject_input_db \
    --faithfulness_eval_db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/test_faith_db \
    --baseline_faithfulness_eval_db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/baseline_faith_db

4. Filter to Create the Final SynthSys Dataset

For the train split:

python scripts/create_synthsys.py \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/train_subject_input_db \
    --save_path $DATA_DIR/$HF_MODEL_NAME/synthsys/train_filtered_data.pkl

For the test split:

python scripts/create_synthsys.py \
    --db_dir $DATA_DIR/$HF_MODEL_NAME/synthsys/test_subject_input_db \
    --save_path $DATA_DIR/$HF_MODEL_NAME/synthsys/test_filtered_data.pkl

SelfDescribe

1. Download Self-Descriptions

aws s3 cp s3://transluce-public/user-modeling/wikipedia_stereotypes.jsonl $DATA_DIR

2. Launch Belief Evaluation and Create the SelfDescribe(8B) Dataset

python scripts/create_self_describe.py \
    --stereotypes_path $DATA_DIR/wikipedia_stereotypes.jsonl \
    --subject_hf_model_id $HF_MODEL_ID \
    --save_dir $DATA_DIR/$HF_MODEL_NAME/self_describe \
    --subject_client_url http://0.0.0.0:$PORT/v1

PRISM

Launch Belief Evaluation and Create the PRISM(8B, Gender) Dataset

python scripts/create_prism.py \
    --save_dir $DATA_DIR/$HF_MODEL_NAME/prism \
    --subject_client_url http://0.0.0.0:$PORT/v1