Skip to content

Releases: decoderesearch/SAELens

v6.14.2

10 Oct 23:17

Choose a tag to compare

v6.14.2 (2025-10-10)

Fix

  • fix: Make run_evals() default behavior match the CLI behavior (#560)

  • Make run_evals() default behavior match the CLI behavior

  • minor tweaks to PR and linting and formatting


Co-authored-by: David Chanin <[email protected]> (c2b0c94)

v6.14.1

10 Oct 19:19

Choose a tag to compare

v6.14.1 (2025-10-10)

Fix

  • fix: notebooks should use cfg.metadata.prepend_bos, not cfg.prepend_bos (#566) (64bcaa8)

v6.14.0

10 Oct 13:35

Choose a tag to compare

v6.14.0 (2025-10-10)

Feature

  • feat: rescale topk SAE decoder during training (#565)

  • feat: rescale topk SAE decoder during training

  • updating docstring for BatchTopKSAEConfig

  • fixing test

  • Update tests/saes/test_topk_sae.py

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

  • Update sae_lens/saes/topk_sae.py

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

  • Update sae_lens/saes/topk_sae.py

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

  • Update tests/saes/test_standard_sae.py

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

  • Update tests/saes/test_topk_sae.py

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

  • Update tests/saes/test_topk_sae.py

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

  • Update tests/saes/test_topk_sae.py

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

  • Update tests/saes/test_batchtopk_sae.py

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

  • Update tests/saes/test_batchtopk_sae.py

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

  • Update tests/saes/test_batchtopk_sae.py

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

  • Update tests/saes/test_batchtopk_sae.py

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

  • Update tests/saes/test_sae.py

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

  • changes from CR

Co-authored-by: Anthony Duong <[email protected]> (e6d8738)

v6.13.1

08 Oct 13:52

Choose a tag to compare

v6.13.1 (2025-10-08)

Chore

  • chore: report 5mb disk usage swings in tests (#563)

  • report 50mb disk usage swings in tests

  • switching to smaller datasets and having smaller threshold for printing disk usage (7ef8946)

  • chore: track disk usage in tests and clean up some space (#562)

  • track disk usage in tests to figure out where disk is going

  • more logging

  • more reporting

  • using older method of outputting disk space in CI

  • output diff in disk space from tests

  • reducing disk usage

  • adding tmate debugging

  • tweaking caches (f3d587f)

Documentation

  • docs: simplifying jumprelu docs (1e556b1)

  • docs: adding tips in the docs for which SAE architectures to use (#556) (382010c)

  • docs: adding docs for making custom SAE classes (#537) (be87864)

Fix

  • fix: Make prepend_bos=True work for disable_concat_sequences case (#559)

  • Make prepend_bos=True work for disable_concat_sequences case

  • fixing formatting

  • more tests and fixed linting


Co-authored-by: David Chanin <[email protected]> (28da94e)

v6.13.0

04 Oct 14:15

Choose a tag to compare

v6.13.0 (2025-10-04)

Chore

  • chore: adding a test for folding scaling factor all architectures (#555) (795363b)

Feature

  • feat: optional sparse-decoding for topK SAEs (#550)

  • Add kernel skeleton and remove redundant b_enc param from TopK class

  • Add option in topK to save SAE activations as a sparse tensor

  • Add sparse activation config flag & update tests

  • Linting changes

  • Making sparse COO tensors compatible with HookedTransformer (WIP)

  • Changes to make sparse SAE intermediate implementation transparent to TransformerLens Hooks

  • Add formatting script (for future optimization)

  • Address PR review comments

  • allow multidim sparsity in topk saes

  • fix logging with sparse feature acts

  • switch TopK to use dense tensors by default in case users are extending / using this module

  • bust CI cache to hopefully get CI to not run out of disk...

  • disable autocase for sparse.mm

  • default TopK SAEs to disable sparse training until we can improve performance

  • updating docs for topk config


Co-authored-by: David Chanin <[email protected]> (3d61660)

v6.12.3

26 Sep 10:00

Choose a tag to compare

v6.12.3 (2025-09-26)

Fix

  • fix: properly convert old SAELens TopK SAEs on load (#554)

  • fix: properly convert old SAELens TopK SAEs on load

  • fixing other loading test (9e71055)

v6.12.2

26 Sep 09:33

Choose a tag to compare

v6.12.2 (2025-09-26)

Fix

  • fix: typo in sae.py leading to misleading warning (#552) (52afbda)

v6.12.1

10 Sep 23:09

Choose a tag to compare

v6.12.1 (2025-09-10)

Fix

  • fix: andy's last name starts with a not r 😓 (3376126)

v6.12.0

10 Sep 22:34

Choose a tag to compare

v6.12.0 (2025-09-10)

Feature

  • feat: add misaligned persona SAEs (#548)

  • feat: add andyrdt llama/qwen SAEs

  • update np names

  • skip redundant/long test (4ec6d52)

v6.11.1

08 Sep 15:06

Choose a tag to compare

v6.11.1 (2025-09-08)

Documentation

  • docs: removing unnecessary fields from sae-table generation (#542) (3432f00)

Fix

  • fix: allow float for batchtopk k param (#547) (c332b6c)