Add SpliceBERT 1024nt results

[chenkenbio]

Summary

• add a GENEB embedding extractor for the original SpliceBERT.1024nt Zenodo checkpoint
• add the harness-generated submission for all 100 tasks under the full, 10-shot, and 1-shot regimes
• add a model card documenting training data, pooling, context handling, and overlap limitations

The encoder is frozen. Sequences longer than the native 1,024-nt context are processed with 1,024-nt windows at a 512-nt stride, with an end-anchored final window and equal averaging of window embeddings. Each window uses final-layer mean pooling over non-padding tokens, including [CLS] and [SEP].

Results

Regime	Micro MCC	Macro MCC
Full	0.5545	0.5680
10-shot	0.3354	0.3402
1-shot	0.1687	0.1545

The mean full-data MCC across the seven splice-site tasks is 0.6334.

Disclosure

• provenance is marked self-reported
• zero-shot status is not claimed because sequence-level pretraining overlap has not been established
• the checkpoint was pretrained on vertebrate primary RNA; potential genomic-locus overlap is documented in the model card
• GENEB sequences shorter than the 64-nt pretraining minimum are evaluated unchanged, while overlength sequences use the documented sliding-window policy

Validation

• python tools/validate_submission.py submissions/splicebert-1024nt.json — valid submission with 100 tasks in each regime
• python tools/build_leaderboard.py — rebuilt successfully with 41 models and 100 tasks across three regimes
• focused extractor tests — 9 passed
• extractor Python compilation and git diff --check — passed
• complete GPU evaluation — 100/100 tasks, exit status 0

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f8a479a007

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".