Library / Framework · Python
event-aware-SAE-trainer
Self-tuning SAE trainer: train a sparse autoencoder on every layer of any model in one unattended run. No per-layer retuning. Using an augmented-Lagrangian controller hits the sparsity target mid-run instead of restarting on failure.
What it is
The repository provides a library for training JumpReLU sparse autoencoders (SAEs) per transformer layer with an event-aware optimizer scheduler and augmented Lagrangian sparsity control. It includes tools for activation capture, checkpointing, and diagnostics, demonstrated by training a 35-layer SAE atlas for Gemma-4-E2B-it.
- activation-steering
- augmented-lagrangian-method
- dictionary-learning
- feature-extraction
- huggingface
- hyperparameter-optimization
- interpretability
- llm
- mechanistic-interpretability
- pytorch
Where it counts
- NLP & LLMs90%
- Machine Learning80%
- MLOps40%
- Computer Vision10%
- Data Science10%
How much of the project is about each field, as scoring counts it. It ranks in the field it is most about.
Around it on the NLP & LLMs board
Whole board- 1ROCmForgeInference System · Rust77.9ExpertLuiz Cesar SpiesNLP & LLMs
- 2regescore-benchmarkBenchmarking Project · Python76.5AdvancedMaster MindNLP & LLMs
- 3event-aware-SAE-trainerLibrary / Framework · Python75.2AdvancedRickNLP & LLMs
- 04RegesMoEML Application · C++72.5AdvancedMaster MindNLP & LLMs
- 05WebsearchRAG System · Python70.9AdvancedJens MittelbachNLP & LLMs
