Are self-explanations from Large Language Models faithful?

This is the code for the paper Are self-explanations from Large Language Models faithful?.

Large language models are increasingly being used by the public, in the form of chat models. These chat systems often provide detailed and highly convincing explanations for their answers, even when not explicitly prompted to do so. This makes users more confident in these models. However, are the explanations true? If not true, this confidence is unsupported which can be dangerous.

We measure the truthfulness (i.e. interpretability-faithfulness) of the explanations that LLMs provide, so called self-explanations. We do so by holding the models accountable to their own explanations, using self-consistency checks. We find that the truthfulness is highly dependent on the model and the specific task. Suggesting we should not have general confidence in these explanations.

install

This module is not published on PyPi but you can install directly with:

python -m pip install -e .

API

The module provides the APIs create custom experiments or reuse existing experiments. As such it's possible to adapt to new models, datasets, or tasks. The API is async.

Docstrings for every class (TGIClient, IMDBDataset, Llama2Model, etc.) are provided in the source files. For a complete example on how to use these class together, please see the ./experiments/analysis.py file.

High level overview of each class

graph TD;
    TGI-->GPU[[e.g. 4x A100 GPUs]];
    VLLM-->GPU[[e.g. 4x A100 GPUs]];
    Cache-->FS[[e.g. Lustre FS]];
    Client--sqlite-->Cache;
    Client-->Offline;
    Client--http-->TGI;
    Client--http/not used-->VLLM;
    Model-->Client;
    Task-->Model;
    Task-->Dataset;
    AsyncMap-->Dataset;
    AsyncMap-->Task;
    Experiment--configuable--->Task;
    Experiment--configuable--->Dataset;
    Experiment-->AsyncMap;
    Experiment--configuable--->Model;
    Experiment--configuable--->Client;

Loading

Short example

import asyncio
import pathlib

from introspect.client import TGIClient
from introspect.dataset import IMDBDataset
from introspect.model import Llama2Model
from introspect.tasks import SentimentCounterfactualTask
from introspect.util import AsyncMap
from introspect.database import GenerationCache

async def main():
    cache = GenerationCache('custom_experiment')
    client = TGIClient('http://127.0.0.1:3000', cache)
    dataset = IMDBDataset(persistent_dir=pathlib.Path('.'))
    model = Llama2Model(client)
    task = SentimentCounterfactualTask(model)
    aggregator = task.make_aggregator()

    async with cache:
        async for answer in AsyncMap(task, dataset.test(), max_tasks=20):
            aggregator.add_answer(answer)

    print(aggregator.results)

if __name__ == '__main__':
    asyncio.run(main())

Experiments

Rather than writing your own code, the experiments from the paper can be run by using the provided CLI: python experiments/analysis.py.

For example:

python experiments/analysis.py
    --persistent-dir $SCRATCH/introspect
    --endpoint http://127.0.0.1:3000
    --task redacted
    --task-config ''
    --model-name llama2-70b
    --dataset IMDB
    --split test
    --seed 0

Arguments

• --persistent-dir controls where data is stored.
• --endpoint is the URL to REST API used for inference. By default the TGI client is used, however you can specify another client with --client.
• --task controls the main experiments, either classify,
• counterfactual, redacted, or importance.
• --task-config allows to use a number of different prompt-variations. Which prompt variation is allowed depends on the task.
  • Classify: c-persona-you, c-persona-human, otherwise objective personal. m-removed for the [REMOVED] token, otherwise [REDACTED].
  • Counterfactual: e-persona-you, e-persona-human, otherwise objective personal. e-implcit-target for the implicit counterfactual target, otherwise explicit is used.
  • Redacted and Importance: e-persona-you, e-persona-human, otherwise objective personal. m-removed for the [REMOVED] token, otherwise [REDACTED].
• --model-name specify the model, either llama2-70b, llama2-7b, falcon-40b, falcon-7b, mistral-v1-7b. This is a shorthand that will resolve to the appropiate huggingface repo and model type. You can also specify these manually with --model-id and --model-type respectively.
• --dataset is which dataset. Included datasets are IMDB, RTE, bAbI-1, MCTest.
• --split either train, valid, or test.
• --seed the seed used for inference.

Running on a HPC setup

For downloading the required resources we provide a experiment/download.py script

Finally, we provide scripts for submitting all jobs to a Slurm queue, in jobs/. The jobs automatically use #SCRATCH/introspect as the persistent dir.