[jmullings]

Adds JMullings/spectral-embed-v2-140m (supersedes v1, which is already in MTEB).

Model: CPU-only English embedding model. Hashed n-gram featurisation (BLAKE2b, Porter stemming + stopwords, BM25-style term weighting, bucket IDF) followed by a learned linear projection to 2048 dimensions. Model card: https://huggingface.co/JMullings/spectral-embed-v2-140m

Changes

• Add SpectralEmbedV2Model / SpectralFeaturizer and the spectral_embed_v2_140m ModelMeta
• Set superseded_by on v1
• Requires nltk for stemming

Loader verification: the loader's output matches the original artifact's served embeddings at ~0.994 mean cosine (n=200). A single serving-only weights file (~537 MB) is used.

Training data: TODO (corpus, teacher model, and whether any labelled/evaluation data was used).

Results: to be submitted in embeddings-benchmark/results#789, regenerated against this implementation and revision 6943d55e.

Draft until the training-data section is complete and the results are regenerated.

[Samoed]

Can't you merge v2 version with v1?

[jmullings]

Good call @Samoed. I've refactored spectral_models.py to merge SpectralEmbedV2Model into SpectralEmbedModel. Both v1 and v2 now use the same unified loader class and share identical encode, batching, and similarity methods, while dispatching to the appropriate featurizer based on the model name. Pushed in the latest commit.

[KennethEnevoldsen]

Won't this make v1 results non-reproducible?

[Samoed]

No, this just adds head to swap between v1 and v2

[KennethEnevoldsen]

my biggest concern it reproducibility of v1.

[KennethEnevoldsen]

remove

[KennethEnevoldsen]

remove

[KennethEnevoldsen]

Is stemming really worth it?

[KennethEnevoldsen]

My experience with it is that is consistently make the model generalize worse, but mostly worked in non-english languages, but would still be surprised if you couldn't do better than stemming.

[jmullings]

Thanks both. On v1 reproducibility (@KennethEnevoldsen):

• The v1 featurization code (normalisation, feature keys, hashing, weights) is unchanged from main. v1 now goes through the shared _embed/encode path, and the v1 ModelMeta differs from main only by superseded_by. Name and revision (39169ba7aa8d1de29da32d97a6129bca790e9193) are unchanged.
• I checked this against the real v1 weights at that revision: embeddings from the unified SpectralEmbedModel are bit-identical to those from main's v1 featurizer (max abs diff 0.0). So existing v1 results should not change.
• One behavioural change: similarity() on two single 1-D vectors now returns a 0-d array instead of a (1, 1) matrix. The values are equal, and 2-D inputs are unchanged. This is for the summarization-style call path.

On stemming: the v2 weights were trained on the Porter-stemmed, stopword-filtered featurization (spectral_v6), so stemming is part of the model rather than an inference option. Removing it would need retraining.

I also removed the banner comments. Re-requesting review from you and @Samoed.

[jmullings]

@Samoed FYI:
On macOS, scikit-learn 1.6.1 produces 0.05358 (passes locally).
On Linux, scikit-learn 1.6.1 produces 0.05445 (fails in CI).

[Samoed]

Yes, I know about that issue, but linux test on lower dependecy is flaky