v6.20.1

v6.20.1 (2025-10-28)

Fix

• fix: update pypi publisher (2541e2c)

v6.20.0

v6.20.0 (2025-10-27)

Feature

• feat: Temporal SAE integration (#575)

• clauded temporal SAE integration

• setting up tests and removing intermediate files

• use SAELens names for weights

• updating implementation

• disallow folding w dec norm for temporal saes

• added temporal sae

• Revert "added temporal sae"

This reverts commit c644790.

• added warning to standalone decoding

• updated loading temporal sae to safetensors format

• fixing syntax issues

• testing sae inference

• fix: 1. W_enc not initialized for tied weights and 2. added scaling factor.

• added end-to-end comparison with original implementation of TemporalSAE

• fixed linting

• fix: set temporal hook_name, fix lint

• add neuronpedia entries to yaml

• use gemma-2-2b instead of google/gemma-2-2b in pretrained yaml

• make W_enc optional for Temporal SAE

• adapted tests

• ruff formatting

• fixed layer index of temporal Llama SAEs

• fix: temporal pretrained yaml

• fix: undo formatting change

• Fix hook_resid_post ID in pretrained_saes.yaml

• fix: final corrections for temporal SAEs llama yaml

• moving scaling into temporal SAEs for now

---

Co-authored-by: David Chanin <[email protected]>
Co-authored-by: Johnny Lin <[email protected]> (888c586)

v6.19.0

v6.19.0 (2025-10-27)

Chore

• chore: update more urls in docs given new repo url (#578) (137e3c7)

• chore: Update image source in README.md (8d8fc97)

• chore: update URLs for new repo (#577) (a8f9adf)

Feature

• feat: adding support for loading goodfire llama 3 SAEs (#579) (f5beb32)

v6.18.0

v6.18.0 (2025-10-22)

Feature

• feat: add LLMRunnerConfig.from_dict() method (#574) (8f16fa5)

v6.17.0

v6.17.0 (2025-10-18)

Feature

• feat: Resume training from checkpoints (#561)

• setting up saving / resuming from checkpoints in training

• adding some tests for checkpointing

• more tests

• adding docs

• reverting notebooks, not sure why they changed

• try busting CI cache

• fixing tests

• fixing tests

• changes from CR (dee399c)

v6.16.3

v6.16.3 (2025-10-18)

Fix

• fix: sync matryoshka implementation with dictionary_learning (#571) (9809d8a)

v6.16.2

v6.16.2 (2025-10-18)

Fix

• fix: improve batchtopk handling of 3d+ inputs (#570) (92ef848)

v6.16.1

v6.16.1 (2025-10-16)

Fix

• fix: pretrained_sae_loader doc file link (#569) (61638d2)

v6.16.0

v6.16.0 (2025-10-14)

Feature

• feat: include step num in train step input (#568) (d7cecc4)

v6.15.0

v6.15.0 (2025-10-13)

Feature

• feat: adding support for training Matryoshka BatchTopK SAEs (#567)

• feat: adding MatryoshkaBatchTopKSAE

• Apply suggestion from @anthonyduong9

Co-authored-by: Anthony Duong <[email protected]>

• changes from CR

• Update sae_lens/llm_sae_training_runner.py

Co-authored-by: Anthony Duong <[email protected]>

---

Co-authored-by: Anthony Duong <[email protected]> (43977ad)