Releases: ApolloResearch/e2e_sae
Releases · ApolloResearch/e2e_sae
v2.1.1
*Fix matplotlib pyright issues
Full Changelog: v2.1.0...v2.1.1
v2.1.0
What's Changed
- Fix autointerp pyright issues by @danbraunai-apollo in #71
- Stop pinning major and minor versions of dependencies
Full Changelog: v2.0.0...v2.1.0
v2.0.0
What's Changed
- Fix dataset tokenization by @danbraunai-apollo in #69
- Divide loss by grad accumulation steps by @danbraunai-apollo in #70
Full Changelog: v1.0.0...v2.0.0
v1.0.0
What's Changed
- Remove dividing loss by n_grad_accumulation_steps by @danbraunai-apollo in #68
Full Changelog: v0.0.3...v1.0.0
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