v6.6.1

v6.6.1 (2025-08-18)

Fix

• fix: compare unscaled input with unscaled output in var explained (#521)

• fix: compare unscaled input with unscaled output in var explained

• move sparsify loading into a closure to avoid doubling memory usage (10d04ae)

• fix: relaxing saelens deps (#519)

• relaxing saelens deps

• fixing types

• fixing tests

• fixing formatting

• pin sparsify version in tests

• hopefully fixing sparsify test

• trying to fix sparsify test another way... (66f575a)

v6.6.0

v6.6.0 (2025-08-14)

Feature

• feat: adding mwhanna qwen transcoders to pretrained list (#515)

• adding mwhanna qwen transcoders to pretrained list

• bust cache to hopefully fix disk space issue in CI

• fix: update transformerlens version, add -lowl0 to pretrained_saes repo_id, add neuronpedia IDs

• fix: test for neuronpedia id on qwen

• adding a mlp.hook_in hook for mwhanna transcoders

• updating model_from_pretrained_kwargs for mwhanna saes

---

Co-authored-by: Johnny Lin <[email protected]> (6e1a04f)

v6.5.3

v6.5.3 (2025-08-11)

Fix

• fix: dont apply b_dec to Gemma Scope transcoder inputs (#516)

• fix: dont apply b_dec to transcoder input

• fixing tests

• change from cr (57888fa)

v6.5.2

v6.5.2 (2025-08-09)

Fix

• fix: allow sae_lens.evals command to work with legacy datasets (#514)

• Fix command

• fixing type issue

• fixing test

---

Co-authored-by: David Chanin <[email protected]> (daeca02)

v6.5.1

v6.5.1 (2025-08-03)

Fix

• fix: Revert "upgrades python-semantic-release to latest (#510)" (#512)

This reverts commit c361af6. (e717cb4)

v6.5.0

v6.5.0 (2025-08-03)

Features

• Load GemmaScope transcoders with SAE.from_pretrained (#509, 72f38ef)

---

Detailed Changes: v6.4.0...v6.5.0

v6.4.0

v6.4.0 (2025-07-31)

Features

• Deprecates eval_hook (#511, 0c91973)

---

Detailed Changes: v6.3.1...v6.4.0

v6.3.1

v6.3.1 (2025-07-26)

Fix

• fix: adds batchtopk to architecture choices (#508) (9c758f7)

Unknown

• upgrades python-semantic-release (#507)

• upgrades python-semantic-release

• downgrades to v9 (596889a)

• deletes TrainSAEOutput (#506) (7ceb4ab)

v6.3.0

v6.3.0 (2025-07-18)

Feature

• feat: adds fields for sequence logic in ActivationsStore (#492)

• adds fields for sequence logic in ActivationsStore

• adds docstrings to LanguageModelSAERunnerConfig

• groups params for ActivationsStore together

• adds disable_concat_sequences to concat_and_batch_sequences()

• adds disable_concat_sequences to PretokenizedDatasetMetadata

• replaces exclude_bos_between_sequences with sequence_separator_token

• updates test name

• fixes tests

• adds params to ActivationsStore.from_sae()

• adds fields to SAEMetdata (25fc5c3)

v6.2.0

v6.2.0 (2025-07-17)

Chore

• chore: updating docs for v6 release (#502) (07196e1)

Feature

• feat: adds aux loss coefficient to TopK training config (#497)

• adds aux loss coefficient to TopK training config

• fixes aux loss coefficient (ce310f5)

Unknown

• Replace allclose with testing assert close (#480)

• replaces torch.allclose() with torch.testing.assert_close()

• removes type: ignore comment

• makes changes to augment generated messages

• extracts repeated code to helpers

• removes unnecessary args

• updates after rebase (f39763f)