Introduction

Delphi was the home of a temple to Phoebus Apollo, which famously had the inscription, 'Know Thyself.' This library lets language models know themselves through automated interpretability.

This library provides utilities for generating and scoring text explanations of sparse autoencoder (SAE) and transcoder features. The explainer and scorer models can be run locally or accessed using API calls via OpenRouter.

The branch used for the article Automatically Interpreting Millions of Features in Large Language Models is the legacy branch article_version, that branch contains the scripts to reproduce our experiments. Note that we're still actively improving the codebase and that the newest version on the main branch could require slightly different usage.

Installation

Install this library as a local editable installation. Run the following command from the delphi directory.

pip install -e .

Getting Started

To run the default pipeline from the command line, use the following command:

python -m delphi meta-llama/Meta-Llama-3-8B EleutherAI/sae-llama-3-8b-32x --explainer_model 'hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4' --dataset_repo 'EleutherAI/fineweb-edu-dedup-10b' --dataset_split 'train[:1%]' --n_tokens 10_000_000 --max_latents 100 --hookpoints layers.5 --filter_bos --name llama-3-8B

This command will:

Cache activations for the first 10 million tokens of the dataset.
Generate explanations for the first 100 features of layer 5 using the specified explainer model.
Score the explanations uses fuzzing and detection scorers.
Log summary metrics including per-scorer F1 scores and confusion matrices, and produce histograms of the scorer classification accuracies.

The pipeline is highly configurable and can also be called programmatically (see the end-to-end test for an example).

To use other scorer types, instantiate a custom pipeline. You can take inspiration from the main pipeline in delphi.__main__.

Caching

The first step to generate explanations is to cache sparse model activations. To do so, load your sparse models into the base model, load the tokens you want to cache the activations from, create a LatentCache object and run it. We recommend caching over at least 10M tokens.

from sparsify.data import chunk_and_tokenize
from delphi.latents import LatentCache

data = load_dataset("EleutherAI/rpj-v2-sample", split="train[:1%]")
tokens = chunk_and_tokenize(data, tokenizer, max_seq_len=256, text_key="raw_content")["input_ids"]

cache = LatentCache(
    model,
    submodule_dict,
    batch_size = 8
)

cache.run(n_tokens = 10_000_000, tokens=tokens)

(See populate_cache in delphi.__main__ for a full example.)

Caching saves .safetensors of dict["activations", "locations", "tokens"].

cache.save_splits(
    n_splits=5,
    save_dir="raw_latents"
)

Safetensors are split into shards over the width of the autoencoder.

Loading Latent Records

The .latents module provides utilities for reconstructing and sampling various statistics for sparse features. The LatentDataset will construct lazy loaded buffers that load activations into memory when called as an iterator object. For ease of use with the autointerp pipeline, we have a constructor and sampler: the constructor defines builds the context windows from the cached activations and tokens, and the sampler divides these contexts into a training and testing set, used to generate explanations and evaluate them.

from delphi.latents import LatentDataset
from delphi.config import SamplerConfig, ConstructorConfig


latent_dict = {
    ".model.layer.0": torch.arange(0, 131072)
}
sampler_cfg = SamplerConfig()
constructor_cfg = ConstructorConfig()

dataset = LatentDataset(
    raw_dir="feature_folder",
    modules=[".model.layer.0"], # This a list of the different caches to load from
    sampler_cfg=sampler_cfg,
    constructor_cfg=constructor_cfg,
    latents=latent_dict,
    tokenizer=tokenizer
)

FAISS Index for Hard Negatives

When constructing features for explanation, you can use FAISS (semantic similarity search) to create hard negative examples. Hard negatives are non-activating examples that are semantically similar to activating examples. This approach:

Creates embeddings for both activating and non-activating examples using the specified embedding model
Builds a FAISS index for efficient similarity search
Finds non-activating examples that are semantically similar to activating examples
Optionally caches embeddings to speed up future runs

To use FAISS for hard negatives, set the non_activating_source parameter to "FAISS" in your ConstructorConfig:

from delphi.config import ConstructorConfig

constructor_cfg = ConstructorConfig(
    non_activating_source="FAISS",
    faiss_embedding_model="sentence-transformers/all-MiniLM-L6-v2",
    faiss_embedding_cache_enabled=True,
    faiss_embedding_cache_dir=".embedding_cache"
)

Contrastive Explainer

The ContrastiveExplainer adds both positive (activating) and negative (non-activating) examples to a single explainer prompt so the explainer model is less likely to label features that are not exclusive to the feature activations (ie we are more likely to provide non activating tokens which are semantically similar). This explainer is automatically used when the non_activating_source is set to "FAISS".

from delphi.explainers import ContrastiveExplainer

explainer = ContrastiveExplainer(
    client,
    threshold=0.3,
    max_examples=15,
    max_non_activating=5,
    verbose=True
)

Generating Explanations

We currently support using OpenRouter's OpenAI compatible API or running locally with VLLM. Define the client you want to use, then create an explainer from the .explainers module.

from delphi.explainers import DefaultExplainer
from delphi.clients import Offline,OpenRouter

# Run locally with VLLM
client = Offline("meta-llama/Meta-Llama-3.1-8B-Instruct", max_memory=0.8, max_model_len=5120, num_gpus=1)

# Run with OpenRouter
client = OpenRouter("meta-llama/Meta-Llama-3.1-8B-Instruct", api_key=key)


explainer = DefaultExplainer(
    client,
    tokenizer = dataset.tokenizer,
)

The explainer should be added to a pipe, which will send the explanation requests to the client. The pipe should have a function that happens after the request is completed, to e.g. save the data, and could also have a function that happens before the request is sent, e.g to transform some of the data.

from delphi.pipeline import process_wrapper

def explainer_postprocess(result):

    with open(f"{explanation_dir}/{result.record.latent}.txt", "wb") as f:
        f.write(orjson.dumps(result.explanation))

    return result

explainer_pipe = process_wrapper(explainer,
    postprocess=explainer_postprocess,
)

The pipe should then be used in a pipeline. Running the pipeline will send requests to the client in batches of paralel requests.

from delphi.pipeline import Pipeline
import asyncio

pipeline = Pipeline(
    loader,
    explainer_pipe,
)

asyncio.run(pipeline.run(n_processes))

Scoring Explanations

The process of running a scorer is similar to that of an explainer. You need to have a client running, and you need to create a Scorer from the '.scorer' module. You can either load the explanations you generated earlier, or generate new ones using the explainer pipe.

RecallScorer(
    client,
    tokenizer=tokenizer,
    batch_size=cfg.batch_size
)

You can then create a pipe to run the scorer. The pipe should have a pre-processer, that takes the results from the previous pipe and a post processor, that saves the scores. An scorer should always be run after a explainer pipe, but the explainer pipe can be used to load saved explanations.

from delphi.scorers import FuzzingScorer, RecallScorer
from delphi.explainers import  explanation_loader,random_explanation_loader


# Because we are running the explainer and scorer separately, we need to add the explanation and extra examples back to the record

def scorer_preprocess(result):
    record = result.record
    record.explanation = result.explanation
    record.extra_examples = record.not_active
    return record

def scorer_postprocess(result, score_dir):
    with open(f"{score_dir}/{result.record.feature}.txt", "wb") as f:
        f.write(orjson.dumps(result.score))

# If one wants to load the explanations they generated earlier
# explainer_pipe = partial(explanation_loader, explanation_dir=EXPLAINER_OUT_DIR)

scorer_pipe = process_wrapper(
        RecallScorer(client, tokenizer=dataset.tokenizer, batch_size=cfg.batch_size),
        preprocess=scorer_preprocess,
        postprocess=partial(scorer_postprocess, score_dir=recall_dir),
    )

It is possible to have more than one scorer per pipe. One could use that to run fuzzing and detection together:

scorer_pipe = Pipe(
    process_wrapper(
        RecallScorer(client, tokenizer=tokenizer, batch_size=cfg.batch_size),
        preprocess=scorer_preprocess,
        postprocess=partial(scorer_postprocess, score_dir=recall_dir),
    ),
    process_wrapper(
        FuzzingScorer(client, tokenizer=tokenizer, batch_size=cfg.batch_size),
        preprocess=scorer_preprocess,
        postprocess=partial(scorer_postprocess, score_dir=fuzz_dir),
    ),
)

Then the pipe should be sent to the pipeline and run:

pipeline = Pipeline(
        loader.load,
        explainer_pipe,
        scorer_pipe,
)

asyncio.run(pipeline.run())

Simulation

To do simulation scoring we forked and modified OpenAIs neuron explainer. The name of the scorer is OpenAISimulator, and it can be run with the same setup as described above.

Surprisal

Surprisal scoring computes the loss over some examples and uses a base model. We don't use VLLM but run the model using the AutoModelForCausalLM wrapper from HuggingFace. The setup is similar as above but for a example check surprisal.py in the experiments folder.

Embedding

Embedding scoring uses a small embedding model through sentence_transformers to embed the examples do retrival. It also does not use VLLM but run the model directly. The setup is similar as above but for a example check embedding.py in the experiments folder.

Scripts

Example scripts can be found in demos. Some of these scripts can be called from the CLI, as seen in examples found in scripts. These baseline scripts should allow anyone to start generating and scoring explanations in any SAE they are interested in. One always needs to first cache the activations of the features of any given SAE, and then generating explanations and scoring them can be done at the same time.

Experiments

The experiments discussed in the blog post were mostly run in a legacy version of this code, which can be found in the Experiments branch.

Development

Set up the pre-commit lint and run the unit tests:

pip install pre-commit pytest
pre-commit install
pytest .

Run an end-to-end test:

python -m delphi.tests.e2e

We use conventional commits for releases.

License

Copyright 2024 the EleutherAI Institute

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

License

kernel-loophole/sae-auto-interp

ErrorLooks like something went wrong!

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Error
Looks like something went wrong!

Packages