stanfordnlp / pyvene

Stanford NLP Python Library for Understanding and Improving PyTorch Models via Interventions

http://pyvene.ai

stanfordnlp/pyvene Issues

[Suggestion]: Support Causal Tracing for LLaMA model
Updated 12 days ago
[P1] Supporting Mistral with Unit Tests (#46)
Closed 14 days ago2
[P1] Multi-GPU model sharding with intervening evaluation and training
Updated 14 days ago
[P0] `IntervenableModel.save()` doesn't save trained model parameters
Closed 14 days ago
[P2] Having a simple tutorial on how to use intervenable model inside another torch module, or pipeline (e.g., HF pipeline)
Closed 14 days ago1
Make docs website
Closed 16 days ago
[P2] Sparse autoencoders
Closed 25 days ago3
[Bug]: RoBERTa giving error with Pyvene 0.1.2. for ReFT. "forward() got an unexpected keyword argument 'use_cache'".
Updated a month ago
[External]: Attention head intervention doesn't work for models other than GPT-2
Closed 2 months ago
[Suggestion]: Docs: explain the first example
Updated 2 months ago
[P0] Making `RepresentationConfig` a PretrainedConfig object
Updated 3 months ago
[Bug]: Whether to support intervene_on_prompt=False
Closed 3 months ago2
[Feature Request / Suggestion]: Capturing of the residual stream
Closed 3 months ago5
[P1] Tuned lens support
Updated 3 months ago
[Bug]: Is this repo support MPT architecture ? i got error
Closed 4 months ago1
[P2] Mac chip MPS mode support
Updated 4 months ago1
[P1] Support more huggingface (transformer-based) models
Updated 4 months ago3
[P0] Adding back GPT2 and other model supports
Closed 4 months ago
[Bug]: pyvene.ai is taking too long to respond
Closed 4 months ago2
[P0] Make interface compatible with HF trainer
Closed 5 months ago
[P2] Add a new huggingface collator for working with Pyvene models
Updated 5 months ago
[P1] Dynamic Intervention Scheduler
Updated 5 months ago
[P1] Adding tests for functions in `modeling_utils.py`
Closed 5 months ago
[P1] Adding tests for interventions and their util functions
Closed 5 months ago
[P2] Enable `use_fast` for head related interventions
Closed 6 months ago1
[P1] Tutorial of Inference-time Intervention
Closed 6 months ago1
[P1] Optionally remove the dependency of the config file
Closed 6 months ago1
[P0] Upgrade tutorials to new API
Closed 6 months ago1
[P1] Speed up training of multiple DAS interventions with caching
Updated 6 months ago
Bug in BoundlessRotatedSpaceIntervention
Closed 6 months ago1
[P1] Supports simple use case of intervention (fixed location intervention for all examples)
Closed 6 months ago
[Bug Fix] `MLPConfig` contains conflict field names with HF that cannot be set to other values when initializing
Closed 6 months ago
[P2] Quantized model support
Updated 6 months ago
[P2] Support interchange intervention training by unfreezing model weights with vanilla intervention
Closed 7 months ago
[P2] Probe training with interventions
Closed 7 months ago2
[P1] Support base-only inputs with activation collection intervention
Closed 7 months ago
[P0] Library Class Renaming Effort
Closed 7 months ago
[P1] Streamlining trainable intervention artifacts saving and sharing
Closed 7 months ago
[P1] Unit tests for main class objects
Closed 7 months ago
[External] Save Interventions
Closed 7 months ago1
[P2] Intervening through time recurrently without teacher-forcing
Updated 7 months ago1
[P0] enable `use_fast` option in the alignable to hyper boost training speed in case intervention locations (for position+subspace) are fixed in a batch
Closed 7 months ago2
[P1] Add unit test for vanilla intervention at different positions and streams
Closed 7 months ago1
[P1] Initial support of stateful model architectures such as RNN
Closed 7 months ago
[P2] Support simple MLP layer for interventions
Closed 7 months ago
[P2] Refactor the AlignableConfig to take in intervention not as type but actual instance
Closed 7 months ago
[P2] Supports list of subspace indices instead of chunk partition
Closed 7 months ago
[P2] Refactor utils in models to smaller files
Closed 7 months ago
[External] Counterfactual sampling in price-tagging tutorial utils code
Closed 7 months ago4
[P1] Introduce grouped interventions where each group is associated with the same source input
Closed 7 months ago