$ the-wire · showcase
SAELens v6.53.0 adds bidirectional AbsTopK; param-decomp expands LM targets
By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology
SAELens shipped a sign-preserving sparse autoencoder variant in v6.53.0, giving operators a new activation mode to train and load against, while param-decomp broadened the language models its CI targets cover.
v6.53.0 jbloomAus/SAELens
v6.53.0 is the release carrying the AbsTopK variant plus a CI fix for HuggingFace rate limits, so a version bump is all it takes to pull both in. If you pin SAELens, this is the tag that contains the new activation.
feat: add AbsTopK bidirectional SAE variant jbloomAus/SAELens
AbsTopK keeps the top-k pre-activations by absolute value with their sign preserved, so latents can fire in either direction instead of only through a ReLU, which is what gives you bidirectional features. It registers alongside TopK and BatchTopK and reuses the existing SAE base classes and TopK machinery, so it slots into your training, eval, and loading paths with no new dependencies.
test: memoize HfApi.list_repo_files to avoid HF rate limits in CI jbloomAus/SAELens
Fork PR runs were dying on HuggingFace 429s because GitHub withholds HF_TOKEN from pull_request runs from forks, and the anonymous limit is per IP and shared across GitHub runners. The fix memoizes HfApi.list_repo_files, which TransformerLens calls on every NeelNanda/* model load such as gelu-1l; if you run SAELens CI from forks, this is what stops the repeated listing calls from exhausting you...
Expand language-model targets, CI architectures, and training diagnostics goodfire-ai/param-decomp
by PD User (shared)
param-decomp adds maintained dense and mixture-of-experts language-model configurations with explicit routed expert selection, plus document-aware inputs, streamed outputs, and router, position, attention, and prompt-targeted diagnostics. It also adds a global transformer CI architecture with depth and site ownership, optionally conditioned on calibrated component activations, with architecture...
