pyvene-interventions

1# pyvene: Causal Interventions for Neural Networks
2 
3pyvene is Stanford NLP's library for performing causal interventions on PyTorch models. It provides a declarative, dict-based framework for activation patching, causal tracing, and interchange intervention training - making intervention experiments reproducible and shareable.
4 
5**GitHub**: [stanfordnlp/pyvene](https://github.com/stanfordnlp/pyvene) (840+ stars)
6**Paper**: [pyvene: A Library for Understanding and Improving PyTorch Models via Interventions](https://aclanthology.org/2024.naacl-demo.16) (NAACL 2024)
7 
8## When to Use pyvene
9 
10**Use pyvene when you need to:**
11- Perform causal tracing (ROME-style localization)
12- Run activation patching experiments
13- Conduct interchange intervention training (IIT)
14- Test causal hypotheses about model components
15- Share/reproduce intervention experiments via HuggingFace
16- Work with any PyTorch architecture (not just transformers)
17 
18**Consider alternatives when:**
19- You need exploratory activation analysis → Use **TransformerLens**
20- You want to train/analyze SAEs → Use **SAELens**
21- You need remote execution on massive models → Use **nnsight**
22- You want lower-level control → Use **nnsight**
23 
24## Installation
25 
26```bash
27pip install pyvene
28```
29 
30Standard import:
31```python
32import pyvene as pv
33```
34 
35## Core Concepts
36 
37### IntervenableModel
38 
39The main class that wraps any PyTorch model with intervention capabilities:
40 
41```python
42import pyvene as pv
43…