v5.6.1

v5.6.1 (2025-03-17)

Fix

• fix: Ensure activation scaling factor set before initializing b_dec when needed (#440)

Co-authored-by: Keltin Grimes <kgrimes@sei.cmu.edu> (1145a62)

• fix: base k_aux on d_in instead of d_sae in topk aux loss (#432)

• fix: base k_aux on d_in instead of d_sae in topk aux loss

• detaching error before aux loss and fixing tests

• changes from cr

• fixing tests (5325c06)

v5.6.0

v5.6.0 (2025-03-12)

Feature

• feat: Add support for dataset name in PretokenizeRunner (#438)

Summary of updates:

• Added optional dataset name parameter
• Updated metadata to track dataset name and column name
• Added test for loading dataset with name argument and checking metadata
• Updated tests to use PretokenizeRunner instead of depricated pretokenizer_runner

Co-authored-by: Keltin Grimes <kgrimes@sei.cmu.edu> (1dc4ef6)

Fix

• fix: fix bug with PretokenizeRunner with streaming=True (#442)

• Fixed issue with PretokenizeRunner with streaming=True

• raise error if streaming with num_proc > 1

---

Co-authored-by: David Chanin <chanindav@gmail.com> (40a5924)

v5.5.2

v5.5.2 (2025-02-25)

Fix

• fix: Merge pull request #435 from jbloomAus/np_r1

fix: Pretrained YAML for ds r1 (871cdf0)

• fix: deepseek (ddb1042)

v5.5.1

v5.5.1 (2025-02-25)

Fix

• fix: Merge pull request #434 from jbloomAus/np_r1

fix: neuronpedia links for DS R1 (cd85f6d)

• fix: neuronpedia links for DS R1 (c335c91)

v5.5.0

v5.5.0 (2025-02-22)

Chore

• chore: skip failing test so CI passes (4f4a7b7)

• chore: fixing docs sae table for deepseek SAE (#421)

• chore: fixing docs sae table for deepseek SAE

• Update sae_lens/toolkit/pretrained_sae_loaders.py

Co-authored-by: Anthony Duong <42191920+anthonyduong9@users.noreply.github.com>

• adding test from CR

---

Co-authored-by: Anthony Duong <42191920+anthonyduong9@users.noreply.github.com> (8760f9e)

Feature

• feat: Merge pull request #431 from Hzfinfdu/main

Supports llama scope r1 distill SAEs (cf754b3)

Fix

• fix: force build (be45fa3)

• fix: force build (a8f38e1)

Unknown

• add all sae paths and meta data to pretrained_saes.yaml (f099d15)

• add all sae paths and meta data to pretrained_saes.yaml (52f0133)

• fix format errors (f99bbbf)

• add support of llamascope R1 distilled (131263d)

v5.4.2

v5.4.2 (2025-02-11)

Chore

• chore: remove ToyModelSAERunnerConfig (#424) (9af8bc0)

• chore: Update config.py to remove duplicate of train_batch_size_tokens: int = 4096 in line 162 and line 188 (#411) (220471b)

Fix

• fix: properly parse CLI dict options as json (#423)

• fix: properly parse CLI dict options as json

• Update config.py (a5ac0f0)

v5.4.1

v5.4.1 (2025-02-03)

Chore

• chore: moves files in tests/unit to tests (#417) (251d3cd)

• chore: removing overlooked legacy file (#416) (02360b0)

• chore: delete legacy files (#415)

• deleting legacy files

• moving tsea stuff into tutorials dir (157e2ad)

Fix

• fix: ensure flattened_mask is on correct device in evals (#418) (38370d4)

v5.4.0

v5.4.0 (2025-02-02)

Feature

• feat: ignore special tokens in SAE training (#413)

• feat: allow masking special tokens during training

• filter out special tokens in activations store before passing to trainer

• ignore special tokens in evals too (9e269df)

v5.3.4

v5.3.4 (2025-01-31)

Chore

• chore: fix sae loading in logits_lens_with_features.ipynb example (#406)

when load sae through SAE.from_pretrained, sae_id should change with layer not always blocks.0.hook_resid_pre (36dab85)

Fix

• fix: remove restrictive typeguard dependency (#414)

merging as this is a problem in the current production release and should be uncontroversial (77de73c)

Unknown

• added NP name to R1 distillation SAE (8ccc556)

• Merge pull request #412 from jbloomAus/feature/qresearch-r1-distill-l19-sae

Added qresearch SAE for distilled R1 LLaMa (32f4a60)

• Update pyproject.toml (84ee3d4)

• several updates to fix typing issues (9482155)

• Added qresearch SAE for distilled R1 LLaMa (08191f4)

v5.3.3

v5.3.3 (2025-01-23)

Fix

• fix: removing invalid gemma scope SAEs from the supported SAEs llist (#408) (60c97f5)