BatuhanECB/FinModernBERT-pairs-sec-synthetic-v1
FinModernBERT-pairs-sec-synthetic-v1 115,238 synthetic finance contrastive-training pairs generated from SEC EDGAR filings — the finance portion of the training set behind FinModernBERT-embed-large-v1, the 395M embedding model that beats the 7B Fin-E5 on FinMTEB Summarization (+0.109) and STS (+0.010). To our knowledge this includes the first public document↔summary positive + mismatched negative pair set built specifically for the FinMTEB Summarization task shape.… See the full description on the dataset page: https://huggingface.co/datasets/BatuhanECB/FinModernBERT-pairs-sec-synthetic-v1.
FinModernBERT-pairs-sec-synthetic-v1
115,238 synthetic finance contrastive-training pairs generated from SEC EDGAR filings — the finance portion of the training set behind FinModernBERT-embed-large-v1, the 395M embedding model that beats the 7B Fin-E5 on FinMTEB Summarization (+0.109) and STS (+0.010).
To our knowledge this includes the first public document↔summary positive + mismatched negative pair set built specifically for the FinMTEB Summarization task shape.
Subsets
from datasets import load_dataset
qp = load_dataset("BatuhanECB/FinModernBERT-pairs-sec-synthetic-v1", "query_passage")
ds = load_dataset("BatuhanECB/FinModernBERT-pairs-sec-synthetic-v1", "doc_summary")Join doc_summary ↔ doc_summary_neg on the query string to build (summary, positivedoc, negativedoc) triplets.
How it was made
- Source text: full-text chunks (~2,000 chars) from SEC EDGAR filings — US-government public-domain material.
- Teacher: an open-weights (Apache-2.0) 35B-A3B MoE instruction model, AWQ-quantized, served with vLLM tensor-parallel on 2× RTX 3090, generating typed pairs per chunk.
- Filtering: faithfulness checks (summary must be entailed by its chunk), near-dup removal, length gates, and a decontamination gate against all FinMTEB-English eval sets (word-shingle overlap; dropped 2,628 dup / 308 length / 81 contaminated / 9 unfaithful). Known-contaminated lineages (FiQA/ConvFinQA etc.) were never used.
- Summary negatives: for each summary, a same-genre chunk from a different filing selected under a positive-aware similarity cap, so negatives are on-topic but wrong.
Reproducing the full training blend (270,238 pairs)
The embed model was trained on this set blended with open general-domain data. We do not rehost those datasets — recipe (sampled with a fixed seed, query/positive schema):
Training recipe (hard-negative mining with intfloat/e5-large-v2 positive-aware filter, CachedMNRL + CoSENT, WiSE-FT) is documented on the model card.
Licensing
- Source chunks: SEC EDGAR filings — public domain (US government works / public regulatory filings).
- Synthetic queries/summaries/scores: generated by an Apache-2.0 teacher, released here under Apache-2.0.
- Note: the separate 5.65B-token DAPT pretraining corpus of the model family is not part of this release and is not releasable (third-party licensing).
Related
- FinModernBERT-embed-large-v1 — the embedding model trained on this data.
- FinModernBERT-large-DAPT — the DAPT foundation.
- FinMTEB — the benchmark (paper).
