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) 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.
Install this library as a local editable installation. Run the following command from the delphi directory.
pip install -e .
To run a minimal pipeline from the command line, you can 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' --max_features 100 --hookpoints layers.5 --dataset_repo 'EleutherAI/rpj-v2-sample' --dataset_split 'train[:1%]'
This will cache the activations of the first 10 million tokens of EleutherAI/rpj-v2-sample, generate explanations for the first 100 features of layer 5 using the explainer model, then score the explanations using fuzzing and detection scorers.
This library uses NNsight to load and edit a model with sparse auxiliary models. We provide wrappers to load GPT-2 autoencoders trained by OpenAI, for the GemmaScope SAEs and for SAEs and transcoders trained by EleutherAI using SAE. See the examples directory for specific examples.
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 FeatureCache object and run it. We recommend caching over at least 10M tokens.
from sparsify.data import chunk_and_tokenize
from delphi.features import FeatureCache
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 = FeatureCache(
model,
submodule_dict,
batch_size = 8
)
cache.run(n_tokens = 10_000_000, tokens=tokens)Caching saves .safetensors of Dict["activations", "locations"].
cache.save_splits(
n_splits=5,
save_dir="raw_latents"
)Safetensors are split into shards over the width of the autoencoder.
The .features module provides utilities for reconstructing and sampling various statistics for sparse features. In this version of the code you needed to specify the width of the autoencoder, the minimum number examples for a feature to be included and the maximum number of examples to include, as well as the number of splits to divide the features into.
from delphi.features import FeatureLoader, FeatureDataset
from delphi.config import FeatureConfig
#
cfg = FeatureConfig(width=131072, min_examples=200, max_examples=10000, n_splits=5)
dataset = FeatureDataset(
raw_dir="feature_folder",
modules=[".model.layer.0"], # This a list of the different caches to load from
cfg=cfg,
)The feature dataset will construct lazy loaded buffers that load activations into memory when called as an iterator object. You can iterate through the dataset using the FeatureLoader object. The feature loader will take in the feature dataset, a constructor and a sampler.
loader = FeatureLoader(
dataset=dataset,
constructor = constructor,
sampler = sampler,
)We have a simple sampler and constructor that take arguments from the ExperimentConfig object. 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.features.constructors import default_constructor
from delphi.features.samplers import sample
from delphi.config import ExperimentConfig
cfg = ExperimentConfig(
n_examples_train=40, # Number of examples shown to the explainer model
n_examples_test=100, # Number of examples shown to the scorer models
n_quantiles=10, # Number of quantiles to divide the data into
example_ctx_len=32, # Length of each example
n_random=100, # Number of non-activating examples shown to the scorer model
train_type="quantiles", # Type of sampler to use for training
test_type="even", # Type of sampler to use for testing
)
constructor = partial(default_constructor, tokens=dataset.tokens, n_random=cfg.n_random, ctx_len=cfg.example_ctx_len, max_examples=cfg.max_examples)
sampler = partial(sample, cfg=cfg)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.feature}.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))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.random_examples
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())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 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 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.
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.
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.
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.