v0.0.3

What's Changed

• Major Refactor by @danbraunai-apollo in #12
• Support more HF datasets by @danbraunai-apollo in #16
• Log discrete metrics such as activation frequency by @danbraunai-apollo in #20
• Tweaks to layerwise-training setup by @danbraunai-apollo in #18
• Create layerwise train script by @danbraunai-apollo in #22
• Add layerwise training by @jordansauce in #17
• Allow forward passes to stop early for efficiency by @jordansauce in #21
• Allow training SAEs on MLP internal activations by @jordansauce in #23
• Generate over-time feature density histograms by @jordansauce in #24
• Use warmup_samples instead of warmup_steps by @danbraunai-apollo in #29
• Use run name prefix only by @danbraunai-apollo in #30
• Load pretrained SAEs into an SAETransformer by @danbraunai-apollo in #26
• Add logit accuracy metrics by @jordansauce in #31
• Add lr schedule with cooldown by @danbraunai-apollo in #32
• Collect output metrics for layerwise runs by @danbraunai-apollo in #35
• Add explained variance by @danbraunai-apollo in #36
• Support uploading multiple splits to HF by @danbraunai-apollo in #39
• Add eval loop and refactor by @danbraunai-apollo in #40
• Calc reconstruction loss at future layers by @danbraunai-apollo in #41
• Avoid altering decoder data attribute for norm by @danbraunai-apollo in #46
• Generative prompt completion by @jordansauce in #47
• Complete support for wandb sweeps by @danbraunai-apollo in #51
• Set min-lr for cosine and linear schedules by @danbraunai-apollo in #53
• Always log future recon loss in eval by @danbraunai-apollo in #52
• Feature Dashboards by @jordansauce in #37
• Add gpt2 configs by @danbraunai-apollo in #55
• Add collect_act_frequency_metrics post-training by @danbraunai-apollo in #57
• Improve documentation and usability by @danbraunai-apollo in #61
• Rename sparsify -> e2e_sae by @danbraunai-apollo in #63
• Make autointerp script by @danbraunai-apollo in #67
• Add plots and feature analysis by @danbraunai-apollo in #56

New Contributors

• @danbraunai-apollo made their first contribution in #12

Full Changelog: https://github.com/ApolloResearch/e2e_sae/commits/v0.0.3