New website for EleutherAI based on Hugo static site generator

Sparse autoencoders

Understanding how features learned by neural networks evolve throughout training

Precisely estimating the volume of basins in neural net parameter space corresponding to interpretable behaviors

Efficient and robust implementation of seq-to-seq automatic piano transcription.

An implementation of model parallel autoregressive transformers on GPUs, based on the Megatron and DeepSpeed libraries

A framework for few-shot evaluation of language models.

Keeping language models honest by directly eliciting knowledge encoded in their activations.