RepoJournal
Interpretability Interpretability
11 wires and counting

$ follow Interpretability

Keep up with Interpretability in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-04
stories 4
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

SAELens 6.53.1 keeps excluded tokens untouched during evals

By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology

kiteretsu903
1 person shipped this
kiteretsu903 @kiteretsu903 2 cited

SAELens shipped a fixed evaluation path for excluded tokens plus the 6.53.1 release that carries it, the only change operators of the library need to know about today.

6.53.1 jbloomAus/SAELens

by semantic-release

Version 6.53.1 is the release that carries the excluded-token fix, so this is the version to pull if you run SAE evaluations. Unchanged otherwise: no config keys, flags, or defaults move.

fix(evals): preserve excluded tokens during SAE reconstruction jbloomAus/SAELens

by kiteretsu903

Evaluation no longer mutates tokens you marked as excluded: with normalization enabled, the residual hook used to keep the scaled activation for those positions, and the attention hooks skipped the exclusion mask entirely. The change is worth checking because a small CPU reproduction with an SAE that simply copies its input shifted a later token's cross-entropy loss from 3.55 to 3.84, which is ...

fix(evals): preserve excluded tokens during SAE reconstruction (#754) jbloomAus/SAELens

by Bozhen Peng

The fix landed as the same preserve-excluded-tokens work merged into the repo, so the correction is on main and in the 6.53.1 tag rather than pending. Operators who pinned an earlier 6.53.x build for evaluation consistency should re-run any comparisons made with excluded tokens set.

$ ls interpretability/reviews/ # the step-back reads

the week in review TransformerLens 4.2.0 fixes Cohere bridge and optimizer parity
Full archive → Weekly reviews →

Quick answers

What shipped in Interpretability on October 4, 2026?
SAELens shipped a fixed evaluation path for excluded tokens plus the 6.53.1 release that carries it, the only change operators of the library need to know about today. In total, 2 commits, 1 pull requests, and 1 releases landed.
Who contributed to Interpretability on October 4, 2026?
1 developer shipped this update, including kiteretsu903.
What were the notable Interpretability updates?
6.53.1, fix(evals): preserve excluded tokens during SAE reconstruction, and fix(evals): preserve excluded tokens during SAE reconstruction (#754).
♥

$ log Interpretability --recent

nnsight 0.8.0 cuts import time from 7.6s to 1.6s 3 shipped nnsight 0.8.0 cuts import time, adds a left-padding fix 7 shipped SAELens 6.54.5 fixes silent feature-density skips 6 shipped Gemma Scope 2 affine transcoders loaded wrongly, SAELens sparsity stats off by one 16 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training verl rejects conflicting vLLM engine_kwargs, open-instruct adds MILES OPD behind a Megatron learner Agent Security Approval digests get RFC 8785 float serialization; sandbox-runtime 0.0.80 kills a 300 ms path

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Alignment & Post-training

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.