Interventions on model-internal states are fundamental operations in many areas of AI, including model editing, steering, robustness, and interpretability. To facilitate such research, we introduce pyvene, an open-source Python library that supports customizable interventions on a range of different PyTorch modules. pyvene supports complex intervention schemes with an intuitive configuration format, and its interventions can be static or include trainable parameters.
Getting Started: 
Since the library is evolving, it is recommended to install pyvene by,
git clone git@github.com:stanfordnlp/pyvene.gitand add pyvene into your system path in python via,
import sys
sys.path.append("<Your Path to Pyvene>")
import pyvene as pvAlternatively, you can do
pip install git+https://github.com/stanfordnlp/pyvene.gitor
pip install pyveneYou can intervene with any HuggingFace model as,
import torch
import pyvene as pv
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "meta-llama/Llama-2-7b-hf" # your HF model name.
model = AutoModelForCausalLM.from_pretrained(
model_name, torch_dtype=torch.bfloat16, device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained(model_name)
def zeroout_intervention_fn(b, s):
b[:,3] = 0. # 3rd position
return b
pv_model = pv.IntervenableModel({
"component": "model.layers[15].mlp.output", # string access
"intervention": zeroout_intervention_fn}, model=model)
# run the intervened forward pass
orig_outputs, intervened_outputs = pv_model(
tokenizer("The capital of Spain is", return_tensors="pt").to('cuda'),
output_original_output=True
)
print(intervened_outputs.logits - orig_outputs.logits)which returns,
tensor([[[ 0.0000, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, ..., 0.0000, 0.0000, 0.0000],
[ 0.4375, 1.0625, 0.3750, ..., -0.1562, 0.4844, 0.2969],
[ 0.0938, 0.1250, 0.1875, ..., 0.2031, 0.0625, 0.2188],
[ 0.0000, -0.0625, -0.0312, ..., 0.0000, 0.0000, -0.0156]]],
device='cuda:0')
Remote Intervention Calls with IntervenableModel on NDIF Backend
We are working with the NDIF team to support remote intervention calls without asking the users to download or host their own LLMs! This is still under construction. All you have to do is to use NDIF library to load your model and use pyvene to wrap it (i.e., pyvene will automatically recognize NDIF models)! Here is an example:
from nnsight import LanguageModel
# load nnsight.LanguageModel as your model to wrap with pyvene
gpt2_ndif = LanguageModel('openai-community/gpt2', device_map='cpu')
tokenizer = AutoTokenizer.from_pretrained('openai-community/gpt2')
# pyvene provides pv.build_intervenable_model as the generic model builder
pv_gpt2_ndif = pv.build_intervenable_model({
"component": "transformer.h[10].attn.attn_dropout.input",
"intervention": pv.CollectIntervention()}, model=gpt2_ndif, remote=False)Then, you can use pv_gpt2_ndif as your regular intervenable model. If you specify remote=True (this is still under construction), then everything will be executed remotely on NDIF server with zero GPU resource required! We provide example code in our main tutorial
The following codeblock can reproduce honest_llama-2 chat from the paper Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. The added activations are only ~0.14MB on disk!
# others can download from huggingface and use it directly
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import pyvene as pv
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-chat-hf",
torch_dtype=torch.bfloat16,
).to("cuda")
pv_model = pv.IntervenableModel.load(
"zhengxuanzenwu/intervenable_honest_llama2_chat_7B", # the activation diff ~0.14MB
model,
)
print("llama-2-chat loaded with interventions:")
q = "What's a cure for insomnia that always works?"
prompt = tokenizer(q, return_tensors="pt").to("cuda")
_, iti_response_shared = pv_model.generate(prompt, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(iti_response_shared[0], skip_special_tokens=True))With this, once you discover some clever intervention schemes, you can share with others quickly without sharing the actual base LMs or the intervention code!
You can also use the pv_gpt2 just like a regular torch model component inside another model, or another pipeline as,
import torch
import torch.nn as nn
from typing import List, Optional, Tuple, Union, Dict
class ModelWithIntervenables(nn.Module):
def __init__(self):
super(ModelWithIntervenables, self).__init__()
self.pv_gpt2 = pv_gpt2
self.relu = nn.ReLU()
self.fc = nn.Linear(768, 1)
# Your other downstream components go here
def forward(
self,
base,
sources: Optional[List] = None,
unit_locations: Optional[Dict] = None,
activations_sources: Optional[Dict] = None,
subspaces: Optional[List] = None,
):
_, counterfactual_x = self.pv_gpt2(
base,
sources,
unit_locations,
activations_sources,
subspaces
)
return self.fc(self.relu(counterfactual_x.last_hidden_state))One key abstraction that pyvene provides is the encapsulation of the intervention schema. While abstraction provides good user-interfact, pyvene can support relatively complex intervention schema. The following helper function generates the schema configuration for path patching on individual attention heads on the output of the OV circuit (i.e., analyzing causal effect of each individual component):
import pyvene as pv
def path_patching_config(
layer, last_layer,
component="head_attention_value_output", unit="h.pos",
):
intervening_component = [
{"layer": layer, "component": component, "unit": unit, "group_key": 0}]
restoring_components = []
if not stream.startswith("mlp_"):
restoring_components += [
{"layer": layer, "component": "mlp_output", "group_key": 1}]
for i in range(layer+1, last_layer):
restoring_components += [
{"layer": i, "component": "attention_output", "group_key": 1}
{"layer": i, "component": "mlp_output", "group_key": 1}
]
intervenable_config = IntervenableConfig(intervening_component + restoring_components)
return intervenable_configthen you can wrap the config generated by this function to a model. And after you have done your intervention, you can share your path patching with others,
_, tokenizer, gpt2 = pv.create_gpt2()
pv_gpt2 = pv.IntervenableModel(
path_patching_config(4, gpt2.config.n_layer),
model=gpt2
)
# saving the path
pv_gpt2.save(
save_directory="./your_gpt2_path/"
)
# loading the path
pv_gpt2 = pv.IntervenableModel.load(
"./tmp/",
model=gpt2)| Level | Tutorial | Run in Colab | Description |
|---|---|---|---|
| Beginner | pyvene 101 | |
Introduce you to the basics of pyvene |
| Intermediate | ROME Causal Tracing | |
Reproduce ROME's Results on Factual Associations with GPT2-XL |
| Intermediate | Intervention v.s. Probing | |
Illustrates how to run trainable interventions and probing with pythia-6.9B |
| Advanced | Trainable Interventions for Causal Abstraction | |
Illustrates how to train an intervention to discover causal mechanisms of a neural model |
Please see our guidelines about how to contribute to this repository.
Pull requests, bug reports, and all other forms of contribution are welcomed and highly encouraged! 
Basic interventions are fun but we cannot make any causal claim systematically. To gain actual interpretability insights, we want to measure the counterfactual behaviors of a model in a data-driven fashion. In other words, if the model responds systematically to your interventions, then you start to associate certain regions in the network with a high-level concept. We also call this alignment search process with model internals.
Here is a more concrete example,
def add_three_numbers(a, b, c):
var_x = a + b
return var_x + cThe function solves a 3-digit sum problem. Let's say, we trained a neural network to solve this problem perfectly. "Can we find the representation of (a + b) in the neural network?". We can use this library to answer this question. Specifically, we can do the following,
- Step 1: Form Interpretability (Alignment) Hypothesis: We hypothesize that a set of neurons N aligns with (a + b).
- Step 2: Counterfactual Testings: If our hypothesis is correct, then swapping neurons N between examples would give us expected counterfactual behaviors. For instance, the values of N for (1+2)+3, when swapping with N for (2+3)+4, the output should be (2+3)+3 or (1+2)+4 depending on the direction of the swap.
- Step 3: Reject Sampling of Hypothesis: Running tests multiple times and aggregating statistics in terms of counterfactual behavior matching. Proposing a new hypothesis based on the results.
To translate the above steps into API calls with the library, it will be a single call,
intervenable.eval_alignment(
train_dataloader=test_dataloader,
compute_metrics=compute_metrics,
inputs_collator=inputs_collator
)where you provide testing data (basically interventional data and the counterfactual behavior you are looking for) along with your metrics functions. The library will try to evaluate the alignment with the intervention you specified in the config.
The alignment searching process outlined above can be tedious when your neural network is large. For a single hypothesized alignment, you basically need to set up different intervention configs targeting different layers and positions to verify your hypothesis. Instead of doing this brute-force search process, you can turn it into an optimization problem which also has other benefits such as distributed alignments.
In its crux, we basically want to train an intervention to have our desired counterfactual behaviors in mind. And if we can indeed train such interventions, we claim that causally informative information should live in the intervening representations! Below, we show one type of trainable intervention models.interventions.RotatedSpaceIntervention as,
class RotatedSpaceIntervention(TrainableIntervention):
"""Intervention in the rotated space."""
def forward(self, base, source):
rotated_base = self.rotate_layer(base)
rotated_source = self.rotate_layer(source)
# interchange
rotated_base[:self.interchange_dim] = rotated_source[:self.interchange_dim]
# inverse base
output = torch.matmul(rotated_base, self.rotate_layer.weight.T)
return outputInstead of activation swapping in the original representation space, we first rotate them, and then do the swap followed by un-rotating the intervened representation. Additionally, we try to use SGD to learn a rotation that lets us produce expected counterfactual behavior. If we can find such rotation, we claim there is an alignment. If the cost is between X and Y.ipynb tutorial covers this with an advanced version of distributed alignment search, Boundless DAS. There are recent works outlining potential limitations of doing a distributed alignment search as well.
You can now also make a single API call to train your intervention,
intervenable.train_alignment(
train_dataloader=train_dataloader,
compute_loss=compute_loss,
compute_metrics=compute_metrics,
inputs_collator=inputs_collator
)where you need to pass in a trainable dataset, and your customized loss and metrics function. The trainable interventions can later be saved on to your disk. You can also use intervenable.evaluate() your interventions in terms of customized objectives.
If you use this repository, please consider to cite our library paper:
@article{wu2024pyvene,
title={pyvene: A Library for Understanding and Improving {P}y{T}orch Models via Interventions},
author={Wu, Zhengxuan and Geiger, Atticus and Arora, Aryaman and Huang, Jing and Wang, Zheng and Noah D. Goodman and Christopher D. Manning and Christopher Potts},
booktitle={arXiv:2403.07809},
url={arxiv.org/abs/2403.07809},
year={2024}
}
If you would like to read more works on this area, here is a list of papers that try to align or discover the causal mechanisms of LLMs.
- Causal Abstractions of Neural Networks: This paper introduces interchange intervention (a.k.a. activation patching or causal scrubbing). It tries to align a causal model with the model's representations.
- Inducing Causal Structure for Interpretable Neural Networks: Interchange intervention training (IIT) induces causal structures into the model's representations.
- Localizing Model Behavior with Path Patching: Path patching (or causal scrubbing) to uncover causal paths in neural model.
- Towards Automated Circuit Discovery for Mechanistic Interpretability: Scalable method to prune out a small set of connections in a neural network that can still complete a task.
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small: Path patching plus posthoc representation study to uncover a circuit that solves the indirect object identification (IOI) task.
- Rigorously Assessing Natural Language Explanations of Neurons: Using causal abstraction to validate neuron explanations released by OpenAI.