Issues

• transformer-lens>=2.16.1 now resolves 4.0.0, which removed transformer_lens.HookedTransformer, so import sae_lens fails at top level

  decoderesearch/SAELens#738 · Filter by author speediedan opened on Sep 22, 2026
  2

• [Proposal] Add hierarchy-conditioned feature-absorption metrics to synthetic SAE evaluation

  decoderesearch/SAELens#729 · Filter by author janmenjayap opened on Aug 30, 2026

• [Bug Report] l0_coefficient has almost no effect on JumpReLU sparsity (Adam normalizes it away)

  decoderesearch/SAELens#719 · Filter by author Hetens opened on Jul 20, 2026
  2

• Tutorial: training an SAE on a speech model (Whisper)

  decoderesearch/SAELens#707 · Filter by author bbrockbrown opened on Jun 21, 2026

• [Proposal] Support data whitening in SAE training

  decoderesearch/SAELens#647 · Filter by author chanind opened on Feb 18, 2026
  3

• [Proposal] Support SwanLab as an alternative logging backend to wandb

  decoderesearch/SAELens#604 · Filter by author wantbook-book opened on Dec 24, 2025
  1

• Is there a tutorial for fine-tuning an SAE?

  decoderesearch/SAELens#584 · Filter by author Chacioc opened on Nov 11, 2025

• [Question] Steer model using multiple features

  decoderesearch/SAELens#564 · Filter by author sn0rkmaiden opened on Oct 8, 2025
  2

• [Question] Canonical SAEs missing matching explanations (Gemma-2-2B)

  decoderesearch/SAELens#549 · Filter by author SleepyLan opened on Sep 19, 2025
  1

• [Question] Why SAETrainer do not normalize W_dec?

  decoderesearch/SAELens#544 · Filter by author jasonrichdarmawan opened on Sep 4, 2025
  8

• [Question] Why SAETrainer do not resample dead neurons?

  decoderesearch/SAELens#543 · Filter by author jasonrichdarmawan opened on Sep 4, 2025
  2

• [Question] Is there any way to train multiple saes at the same time?

  decoderesearch/SAELens#540 · Filter by author neilwen987 opened on Sep 2, 2025
  1

• [Question] Why GatedTrainingSAE do not use a detached forward pass to compute the auxiliary loss?

  decoderesearch/SAELens#539 · Filter by author jasonrichdarmawan opened on Sep 1, 2025
  4

• [Proposal] "Recommended/Featured SAEs" in Docs

  documentation
  decoderesearch/SAELens#530 · Filter by author hijohnnylin opened on Aug 23, 2025

• [Proposal] Documentation + Notebook for each SAE/coder type

  decoderesearch/SAELens#498 · Filter by author hijohnnylin opened on Jun 28, 2025

• [Question] Does SAELens currently support distributed training across multi nodes?

  decoderesearch/SAELens#491 · Filter by author Geaming2002 opened on Jun 5, 2025

• [Question] OOM on A800 80G while train gemma-2-2b SAEs

  decoderesearch/SAELens#490 · Filter by author hermitgreen opened on Jun 5, 2025
  2

• [Bug Report] Huggingface and HookedSAETransformer noticably different

  decoderesearch/SAELens#469 · Filter by author AMindToThink opened on May 1, 2025
  3

• [Proposal] Adding test data metrics throughout the training process

  decoderesearch/SAELens#446 · Filter by author Noam-Diamant opened on Mar 15, 2025
  1

• [Question] How to get correct pipeline for SAE?

  decoderesearch/SAELens#445 · Filter by author EoinKenny opened on Mar 14, 2025

• Feature Request: Support custom hf_model in config / loaders

  decoderesearch/SAELens#436 · Filter by author hijohnnylin opened on Feb 25, 2025

• [Proposal] Optimize decoding in training using torch.nn.functional.embedding_bag

  decoderesearch/SAELens#428 · Filter by author chanind opened on Feb 16, 2025
  1

• [Proposal] Document loading existing SAEs for fine-tuning

  decoderesearch/SAELens#426 · Filter by author chanind opened on Feb 11, 2025
  1

• [Question] High MSE Loss in Released Gemma SAE Model – Am I Using It Correctly?

  decoderesearch/SAELens#410 · Filter by author yingjiahao14 opened on Jan 28, 2025
  2

• [Proposal] Upload Steered Models to Huggingface Hub

  decoderesearch/SAELens#407 · Filter by author AMindToThink opened on Jan 18, 2025