This repository contains the official implementation of the Inverse Constitutional AI (ICAI) algorithm [paper]. ICAI compresses pairwise preference datasets into a readable list of principles (constitution) that the annotations appear to follow (e.g. "select the friendlier response"). ICAI principles provide an interpretable overview of a feedback dataset, enabling users to discover problematic annotation biases or better understand differences between datasets, user groups or models.
- Pip install the package (for development installation see here)
pip install git+https://github.com/rdnfn/icai.git - Set up API secrets: inside the main directory of the cloned repo (or wherever you like really) set up a
secrets.tomlfile like below. You only need to include keys for APIs you want to use.OPENAI_API_KEY="<YOURKEY>" ANTHROPIC_API_KEY="<YOURKEY>"
You can run your first Inverse Constitutional AI (ICAI) experiment using the icai-exp command:
icai-exp data_path="data/processed/example/example.csv"
This will run the ICAI algorithm on the toy example.csv pairwise feedback dataset and generate a constitution for this dataset.
To get the available experiment parameters and instructions on how to adjust them, run
icai-exp --help
Note
If you want more control:
the icai-exp command executes the run function inside ./src/inverse_cai/experiment/core.py. Edit that file, and the other parts of the inverse_cai Python package it uses, to fully adapt this code to your needs.
By default all experiment results are saved in the ./outputs/<DATE>_<TIME> directory. The exact result file location is also printed to the console at the end of any (completed) icai-exp call. These outputs contain a full record of API calls, as well as intermediate states of the algorithm (proposed principles, clusters, distilled principles, etc.). Each result output follows the structure below:
./outputs
└── 2024-03-14
└── 16-09-43
├── api_calls.jsonl
├── core.log
├── main.log
└── results
├── 010_principles_per_comparison.csv
├── 011_principles_list.json
├── 020_principle_clusters.json
├── 030_distilled_principles_per_cluster.json
├── 040_votes_per_comparison.csv
├── 041_votes_per_cluster.json
├── 050_filtered_principles.json
├── 060_constitution.json
├── 092_results_training.csv
└── 093_results_testset.json
To run ICAI on your dataset, you first need to convert it to a csv file with the following three columns: text_a,text_b,preferred_text. The first two should be strings. Note that the ICAI implementation currently uses no separate "prompt" column. If such a column exists in your dataset, you likely want to add the prompt to each response column (text_a,text_b) such that the ICAI algorithm can understand the full context of the preference label. Entries in the column preferred_text should take one of two values: "text_a" or "text_b". Ties or other annotation values are currently not used by the algorithm. To run ICAI on you prepared dataset, simply use:
icai-exp data_path="<path/to/your-data.csv>"
In the exp/configs folder there is a number of configs to recreate experiments. You can run these experiments using the command:
icai-exp -cd ./exp/configs/<EXPERIMENT-DIR>
For example:
icai-exp -cd ./exp/configs/001_synthetic_orthogonal
Note
To re-run paper experiments:
Look at the README file inside the exp/configs. This file gives detailed instructions on which configurations to run, and how to generate the corresponding plots.
clone repo locally, e.g.
git clone git@github.com:rdnfn/icai.git
Then (inside repo folder) install package in editable mode:
pip install -e .
Tests are included as part of the package. Run them using:
pytest ./srcThis doesn't do any meaningful experimental work but allows running the experiment script for testing purposes.
icai-exp generate_constitution=false annotator.constitution=null annotator.other_annotator_configs="[]"
Feedback data plays an important role in fine-tuning and evaluating state-of-the-art AI models. Often pairwise text preferences are used: given two texts, human (or AI) annotators select the “better” one. Such feedback data is widely used to align models to human preferences (e.g., reinforcement learning from human feedback), or to rank models according to human preferences (e.g., Chatbot Arena). Despite its wide-spread use, prior work has demonstrated that human-annotated pairwise text preference data often exhibits unintended biases. For example, human annotators have been shown to prefer assertive over truthful texts in certain contexts. Models trained or evaluated on this data may implicitly encode these biases in a manner hard to identify. To be able to better understand existing pairwise text preference data, we formulate its interpretation as a compression task: the Inverse Constitutional AI problem. Read the full paper for more background.
The figure below provides an overview of the Inverse Constitutional AI (ICAI) problem we introduce: starting from a set of pairwise preference feedback, we derive a set of natural language principles (a constitution) that explain the preference data. For validation, we re-construct the original preferences with an LLM judging according to the generated constitution. The constitution represents a (highly compact) compression of the preferences.
We introduce a first Inverse Constitutional AI (ICAI) algorithm that generates a set of principles based on a feedback dataset. See the figure below for an overview of the algorithm. Given a dataset of pairwise rankings, in Step 1 candidate principles are generated using an LLM. In Step 2, these principles are clustered using an embedding model. In Step 3, similar principles are “de-duplicated” by sampling one principle per cluster. In Step 4, each principle is tested to evaluate its ability to help an LLM reconstruct the original annotations. Finally in Step 5, the principles are filtered according to the testing results and a set of filtered principles are returned as the final constitution. Optionally, this last step is augmented with additional clustering and subsampling steps to ensure diverse principles. The implementation is provided in this repository.
All code in this repo is licensed under Apache-2.0.