Development of and interpretability using disentangled architectures.
Contains code for training sparse feedforward transformers, regular transformers, SAEs, and Transcoders.
Train GptNeoX-style base transformers with either sparse or dense feedforwards:
python -m clearnets.train.train_transformer
python -m clearnets.train.train_transformer --dense
python -m clearnets.train.train_transformer --dense --dataset "lennart-finke/SimpleStories" --tokenizer "EleutherAI/gpt-neo-125m"
sparse_gptneox and sparse_gptneox_config are both modified from the GptNeoX implementation in HuggingFace transformers with a flag to enable training with sparse feedforwards.
Train a comparison SAE or transcoder on a dense feedforward transformer using:
python -m clearnets.train.train_sae_cli --run_name Transcoder-roneneldan--TinyStories-8M --ckpt "data/roneneldan--TinyStories/Dense-TinyStories8M-s=42-full-vocab/checkpoints/last.ckpt" --hookpoints transformer.h.*.mlp --ctx_len 512 --batch_size 128 --grad_acc_steps 2 --max_examples 768_000 --transcode
Run:
python -m clearnets.autointerp.e2e_clearnet
with the following config arguments set to match the experiment:
- model
- sparse_model
- hookpoints
- mlp_mode
Experiments for tracking generalization of neural networks using disentangled activations.