METR Faithfulness and Monitorability Experiments

This is an evaluation designed to test under what circumstances LLMs' CoTs might be faithful and monitorable.

Usage

To collect data on clue difficulty:

python src/main_difficulty.py

Args relevant to this script:

• --model: The model to test faithfulness on (default: anthropic/claude-3-7-sonnet-latest)
• --base_model: The model to test difficulty without reasoning (default: anthropic/claude-3-7-sonnet-latest)
• --temperature: The model temperature when getting difficulty and faithfulness. Default to 0.6. Temperature doesn't affect claude extended thinking mode.
• --batch_size: The number of queries to send in parallel leveraging batch API.
• --max_connections: The maximum number of connections to make with the provider. If batch_size is provided, this will be set to equal batch_size.
• --max_tasks: The number of tasks to run in parallel in an eval set. Default to 5.

To collect data on faithfulness or monitorability:

python src/main_faithfulness.py

Args relevant to this script:

• --model: The model to test faithfulness on (default: anthropic/claude-3-7-sonnet-latest)
• --base_model: The model to test difficulty without reasoning (default: anthropic/claude-3-7-sonnet-latest)
• --temperature: The model temperature when getting difficulty and faithfulness. Default to 0.6. Temperature doesn't affect claude extended thinking mode.
• --batch_size: The number of queries to send in parallel leveraging batch API. Not including this arg will mean that we're not using the batch api. Recommended batch_size is 2000-5000.
• --max_connections: The maximum number of connections to make with the provider. If batch_size is provided, this will be set to equal batch_size.
• --max_tasks: The number of tasks to run in parallel in an eval set. Default to 5.
• --filtered_csv: The filtered out questions that the model cannot solve
• --question_prompt: The question prompt used to introduce the hard problem and the hint.
• --judge_prompt: The LLM Judge prompt used to decide whether the CoT is faithful / monitorable
• --score_faithfulness: A store_true flag indicating this is a faithfulness judge, which does see the hint provided to the model

To generate graphs:

python graph.py

By default, this script goes through the faithfulness and monitorability folder under results/ and create graphs for all the JSON files under the folder, excluding the ones under subset/ folders. Args: --file: You can optionally pass in a result json path to only create the graph for that particular result file. --hint_taking_threshold: remove clues that has a hint taking rate below the threshold from the propensity graph. Default to 0.1

Usage OUTDATED

The entry point is src/main.py.

python src/main.py

The file takes a few arguments:

• --model: The model to test faithfulness on (default: anthropic/claude-3-7-sonnet-latest)

• --base_model: The model to test difficulty without reasoning (default: anthropic/claude-3-7-sonnet-latest)

• --filtered_csv: Path to the csv file with difficulty scores instead of re-running filtering. Default to None.

• --redteam: Run evaluation on a subset of test cases, stores true.

• --temperature: The model temperature when getting difficulty and faithfulness. Default to 0.6, which Qwq's default temperature. Claude 3.7 Sonnet's default temperature is 1

• --max_connections: Number of concurrent request to send to the providers. Default to 20. One might want to customize this based on the rate limits.

• --question_prompt: Path to the question prompt file. Default to utils/prompts/default.py.

• --judge_prompt: Path to the judge prompt file. Default to utils/judge_prompts/faithfulness_0611.txt.

• --score_faithfulness: Whether to score faithfulness. Default to false.

Legacy / Compatibility arguments:

• --display: Display mode for evaluation progress (default: none, other options are full) (broken)
• --testing_scheme: Testing scheme to use, one of: base, rl_no_reasoning, rl_dots (default: base)

Output

By default, the outputs are two graphs and a log of the run in the logs directory.