jbloomAus/SAELens

Issues

• [Proposal] Move to Ruff for linting / formatting

  #368 opened 3 months ago by chanind
  2

• [Proposal] Implement OpenAI's TopK Auxilliary for preventing dead latents.

  #202 opened 3 months ago by jbloomAus
  0

• Use Huggingface Datasets for ActivationStore

  #320 opened 3 months ago by tom-pollak
  0

• [Proposal] Rename `l1_coefficient` to `sparsity_coefficient`

  #360 opened 3 months ago by chanind
  1

• [Bug Report] scale_sparsity_penalty_by_decoder_norm option is ignored

  #363 opened 3 months ago by chanind
  0

• I have checked that there is no similar issue in the repo (required)

  #364 opened 3 months ago by chanind
  1

• [Bug Report] get_sae_config doesn't return the same config as SAE.from_pretrained due to defaulting

  #351 opened 3 months ago by hijohnnylin
  2

• [Proposal] Support training JumpReLU SAEs

  #330 opened 3 months ago by chanind
  4

• Training time

  #353 opened 4 months ago by merlinarer
  1

• [Bug Report] fold_activation_norm_scaling_factor missing division of decoder bias

  #354 opened 4 months ago by tuomaso
  0

• [Proposal] Allow excluding special tokens when applying SAEs in HookedSAETransformer

  #350 opened 4 months ago by chanind
  1

• [Bug Report]

  #337 opened 4 months ago by Yoon-Jeong-ho
  6

• [Bug Report] Error Term flag prevents Feature Ablation in Gemma 2 2b

  #326 opened 4 months ago by gboxo
  2

• [Proposal] Add test coverage to GatedSAE and JumpReLU implementations

  #323 opened 4 months ago by chanind
  0

• [Bug Report] read_sae_from_disk should default to the SAE dtype unless it's overridden

  #230 opened 4 months ago by hijohnnylin
  3

• [Proposal] Either explicitly check notebook formatting in CI, or exclude notebooks from make format

  #287 opened 4 months ago by chanind
  2

• [Proposal] BOS should be ignored by default in the activation store for SAEs not trained on the bos (and future SAEs trained in SAE Lens)

  #333 opened 4 months ago by jbloomAus
  2

• [Proposal] Move files in `tests/unit` to `tests`

  #335 opened 4 months ago by anthonyduong9
  1

• [Proposal] Refactor `generate_sae_table.py` and `pretrained_sae_loaders.py` to share code

  #305 opened 4 months ago by chanind
  2

• [Proposal] Simple get_sae_config method for pretrained SAEs

  #315 opened 4 months ago by hijohnnylin
  5

• [Bug Report] When I‘m preparing SAE: ’non-default argument 'l1_coefficient' follows default argument

  #313 opened 4 months ago by YokeYao
  3

• [Proposal] Make the SAE Lens API docs more comprehensive.

  #319 opened 4 months ago by jbloomAus
  0

• [Bug Report] SAE training tutorial metrics do not match linked run

  #276 opened 5 months ago by naterush
  3

• [Bug Report] embedding layer models missing from canonical

  #316 opened 4 months ago by dribnet
  0

• [Bug Report] Unable to train Mamba SAE

  #311 opened 5 months ago by joelburget
  2

• [Question] Llama-3-8B-it-res model not showing up in the pretrained_saes.yaml file

  #302 opened 5 months ago by CallterC
  3

• [Proposal] Simplify SAE IDs for Gemma Scope Canonical SAEs

  #289 opened 5 months ago by ArthurConmy
  2

• [Proposal] Allow users to shuffle ActivationCache dataset, rather than shuffling the pieces of the activation cache

  #277 opened 5 months ago by naterush
  2

• [Proposal] Use stop_at_layer when running with cache in evals

  #245 opened 5 months ago by ckkissane
  1

• [Proposal] Support Remaining GPT2 Small OpenAI SAEs in SAE Lens

  #204 opened 5 months ago by jbloomAus
  0

• [Bug Report] `make unit-test` fails

  #298 opened 5 months ago by niniack
  2

• [Question] How to keep the consistency of ActivationStore

  #297 opened 5 months ago by luckyeric320
  0

• [Bug Report] When running ActivationsStore get error FileNotFoundError: Dataset Skylion007/openwebtext is not cached in None

  #284 opened 5 months ago by MClarke1991
  1

• [Bug Report] CacheActivationsRunner KeyError on Hookpoints PHi3-mini-instruct model.

  #260 opened 6 months ago by prof-schacht
  1

• [Proposal] Download tracking, code snippets, and filtering on the Hugging Face Hub

  #234 opened 6 months ago by osanseviero
  5

• [Bug Report] Slack link in documentation does not work

  #253 opened 6 months ago by thedch
  0

• [Question] Clarification on Feature Annotation in SAEs

  #248 opened 6 months ago by danjuan-77
  1

• [Bug Report] sae-lens and umap-learn compatibility issue

  #246 opened 7 months ago by wlg1
  4

• [Proposal] Add use_error_term as an optional param for HookedSAETransformer.add_sae

  #231 opened 7 months ago by neelnanda-io
  1

• [Proposal] Make SAE Lens compatible with the latest Transformers / TransformerLens versions

  #209 opened 7 months ago by jbloomAus
  0

• [Proposal] Save norm scaling factor

  #222 opened 7 months ago by ckkissane
  1

• [Bug Report] Fine-tuning SAEs with from_pretrained_path seems broken

  #227 opened 7 months ago by ckkissane
  0

• [Proposal] Add a custom SAE.to method to update sae.dtype and sae.cfg.dtype when running sae.to(torch.float32)

  #233 opened 7 months ago by neelnanda-io
  0

• [Bug Report] hook_sae_input should not be on x - self.b_dec, it should be on x

  #232 opened 7 months ago by neelnanda-io
  1

• [Bug Report] Backwards hooks broken in v2.x.x

  #228 opened 7 months ago by chanind
  1

• [Bug Report] load_from_pretrained should respect device parameter

  #220 opened 7 months ago by hijohnnylin
  0

• [Bug Report] Model saving fails on models with `/` in the name

  #223 opened 7 months ago by robertzk
  0

• [Bug Report] Not able to load pretrained SAE using gated architecture

  #208 opened 7 months ago by JThh
  2

• [Bug Report] When installing sae_lens get warning that "cannot import name 'deprecated' from 'typing_extensions'" when importing

  #203 opened 7 months ago by MClarke1991
  1

• [Bug Report] Log sparsity artifact seems wrong

  #197 opened 8 months ago by ckkissane
  2