Spaces:

---

FinanceMTEB

/

FinMTEB

like 10

Sleeping

Hello! Submitting a new open-weights model for the English leaderboard (also emailing per the Submit-tab instructions).

I would like to submit a new model to the FinMTEB (English) leaderboard.

1. Model name and link

• Model: FinModernBERT-embed-large (395M parameters, ModernBERT-large based)
• Link: https://huggingface.co/BatuhanECB/FinModernBERT-embed-large
• Open weights, Apache-2.0.

2. Per-dataset scores (all 35 English tasks, main metric per task)

Evaluated with the official FinMTEB harness (github.com/yixuantt/FinMTEB), default task
main metrics (nDCG@10 retrieval, Spearman STS/Summarization, MAP reranking,
accuracy / v-measure / AP for classification / clustering / pair-classification).

Per-type means: STS 0.4444 · Summarization 0.5883 · Retrieval 0.5021 · Reranking 0.9612 ·
Classification 0.6233 · Clustering 0.5131 · PairClassification 0.6172 · Overall 0.6071.

3. Methodology notes

• Architecture: single ModernBERT-large encoder (395M), mean pooling, cosine
  similarity, 1024-dim embeddings, max_seq 512. Asymmetric query: / passage:
  text prefixes.
• Long documents: inputs over the window are embedded by chunking into ≤16 windows
  of ~506 body tokens, embedding each, L2-normalizing, mean-pooling, re-normalizing.
• Training pipeline: (1) domain-adaptive MLM on 5.65B tokens of finance text
  (decontaminated against all FinMTEB-EN eval sets via word-shingle overlap before
  training; eval-source datasets and contaminated lineages excluded); (2) contrastive
  training on 270,238 pairs (59% finance; synthetic SEC-derived query→passage, graded
  STS, doc↔summary positives + mismatched score-0 negatives, term→definition; blended
  with open general-domain data), InfoNCE (CachedMNRL, scale 20) + CoSENT, positive-aware
  hard negatives mined with e5-large-v2; (3) WiSE-FT weight interpolation
  (0.65·contrastive + 0.35·DAPT) as the released checkpoint.
• DAPT foundation also released: https://huggingface.co/BatuhanECB/FinModernBERT-large-DAPT

Thank you — happy to provide raw result JSONs or re-run details on request.

Best regards,
Batuhan (BatuhanECB)