Generalization Analogies: A Testbed for Generalizing AI Oversight to Hard-To-Measure Domains

Read our paper here. Check out our website where you can browse samples from our datasets here.

Abstract

As AI systems become more capable and are deployed in complex environments, it may become challenging to verify that they follow instructions; however, the limitations of human oversight could be overcome by controlling how LLMs generalize human feedback to contexts where it is unreliable. To better understand how Reward Models generalize human feedback, we craft 69 distribution shifts spanning 8 different categories. We find that Reward Models do not learn to evaluate instruction-following by default and instead favor personas that resemble internet text. Techniques for interpreting Reward Model’s internal representations achieve better generalization, but still frequently fail to distinguish instruction-following from conflated behaviors. We consolidate the 15 most challenging distribution shifts into the \textbf{GEN}aralization analog\textbf{IES} (\textsc{GENIES}) benchmark, which we hope will enable progress toward controlling Reward Model generalization.

Quickstart

This repository contains:

• Our datasets (./distributions) along with pairing specifications (./distribution_shifts).
• Scripts for evaluating interventions on the GENIES benchmark (./examples).
• Our results (./results).
• Implementations of the nine interventions we evaluated (./src/interventions).

All of the models we fine-tuned with Lora can be found on huggingface.

Setup:

conda create --name env python=3.10
pip install -e .
python download_model_from_hf.py EleutherAI/pythia-410m models/pythia-410m

WARNING: pythia-410m is mostly useful for testing purposes. Most tuning interventions perform poorly with this model.

APIs

The primary api is api/compute_generalization_metrics, which receives a base model, intervention directory, and a collection of distribution shifts, and computes various generalization metrics. See examples/compute_generalization_metrics.sh for example usage.

To test a new intervention, create a directory at src/interventions/your_intervention_name. This directory must contain a train.py file and an eval.py file.

src/interventions/your_intervention_name/train.py should be a script that accepts the following arguments:

• model_dir (str): the directory of the base model that is being trained.
• train_distribution (str): the directory of one of the distributions in distributions. For example: distributions/alpaca_mmlu.
• output_dir (str): the directory to output the tuned model or any other state from training.

src/interventions/your_intervention_name/eval.py should be a script that accepts the following arguments:

• model_dir (str): the directory of the trained model.
• distribution_dirs (List[str]): a list of subdirectories of distributions.
• output_paths (List[str]): where to save the results. The results should be json files. The only required key is eval_accuracy. Evaluation results are stored in results/evaluations.# GENIES