SAE: wire up MiniLM Stage-J 49K (Matryoshka) end to end

[enjalot]

What

Makes the newly published enjalot/sae-all-MiniLM-L6-v2-stagej-49K (Matryoshka, nested 512/2048/8192/49152, k=64) usable in latent-scope for any dataset embedded with all-MiniLM-L6-v2, and fixes two latent bugs found along the way.

Changes

• web/src/lib/SAE.js — live registry entry for the MiniLM Stage-J SAE (64_128, label parquet served by the latent-taxonomy site, same pattern as nomic).
• web/src/contexts/ScopeContext.jsx — Explore previously looked up SAEs by exact saeAvailable[model_id] while Setup used the prefix-tolerant getSaeForModel(). A transformers-/huggingface--prefixed embedding id could pass Setup and then silently load no features in Explore. Both paths now use getSaeForModel(). Also exposes the matched saeEntry in context.
• web/src/components/Explore/FeatureModal.jsx — the latent-taxonomy deep link was hardcoded to NOMIC_FWEDU_25k; it now uses the scope's SAE registry label (falls back to nomic).
• pyproject.toml — latentsae>=0.2.0,<1; 0.1.3 cannot load Matryoshka checkpoints (cfg.json carries matryoshka_* fields).
• latentscope/scripts/sae.py — stale default repo id …-10BT → …-100BT (matches the registry and the actually-published repo).
• Test mocks for ../lib/SAE updated to export getSaeForModel.

Testing

• npm test: 103/103 pass (the two context suites failed before the mock update, all green after).
• npm run production: builds clean; npm run lint: 0 errors (49 pre-existing warnings, none introduced).
• Python path verified against the public HF repo with latentsae 0.2.0 (installed from git): Sae.load_from_hub('enjalot/sae-all-MiniLM-L6-v2-stagej-49K', '64_128', 'cpu') → num_latents=49152, encode() returns [batch, 64] top_acts/top_indices — exactly the calls ls-sae makes. The ColBERT token repo (32_512) also loads and encodes with the same package version (L0 ≈ 32).

Notes

• The label parquet URL (…/models/MINILM_STAGEJ_49K/features.parquet) goes live when the latent-taxonomy multi-model site deploys; until then Explore fetches 404 for MiniLM scopes (nomic unaffected). The parquet's column order matches the positional parsing in apiService.getFeatures (verified: identical schema to the nomic file).
• web/src/components/FilterDataTable.jsx (root, not Explore/) contains a stale copy of FeatureModal with the same hardcoded link but appears to be imported nowhere; left untouched.
• Per-token SAEs (the ColBERT model) are deliberately not wired up here — the SAE pipeline is one-vector-per-row end to end and token-level features are a separate design effort.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3a45c221e7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Update uv.lock with the latentsae floor

This raises the runtime requirement to latentsae>=0.2.0 for the new Matryoshka SAE path, but the committed uv.lock was not regenerated: it still pins the latentsae package to 0.1.3 and still records the project dependency as >=0.1.3,<1 (uv.lock lines 2947-2973 and 3048). Any environment installed from the committed lock will miss the Matryoshka support that this commit relies on, so ls-sae for the new MiniLM entry can fail despite pyproject.toml advertising the newer dependency.

Useful? React with 👍 / 👎.

[enjalot]

Updates:

• Rebased onto current main (the UI revamp + Amber Console restyle merged after this branch was cut). All changes carried over cleanly — the restyled FeatureModal keeps the parameterized taxonomy link, and apiService.getFeatures still parses the label parquet positionally, so the registry URL assumption holds post-revamp.
• Both review comments addressed:
  • uv.lock regenerated — latentsae 0.2.0 is now on PyPI (https://pypi.org/project/latentsae/0.2.0/), so the lock resolves the new floor.
  • Setup/Sae.jsx no longer blindly attaches saes[0]: existing runs are filtered by embedding_id and SAE model_id before selection, in both the initial fetch and the post-job refresh, so a nomic run can't be attached to a MiniLM scope now that the registry has two entries.
• Heads-up: vite build is broken on main itself (MembersTooltip.module.scss was never committed; vitest doesn't catch it because CSS modules are mocked). Fix in Fix build: add missing MembersTooltip.module.scss #161 — with that commit applied on top, this branch builds clean. Tests: 103/103, lint: 0 errors.