gte-large-formbench-mnrl

A domain-adapted sentence-transformers model derived from Alibaba-NLP/gte-large-en-v1.5 and fine-tuned on the FormBench retrieval benchmark for formulation chemistry. It maps passages from formulation patents into a 1024-dimensional dense vector space and is optimised for within-domain retrieval among structurally similar near-miss passages — the central capability targeted by FormBench.

This repository hosts an anonymised release for NeurIPS 2026 double-blind review.

Model details

Item	Value
Base model	Alibaba-NLP/gte-large-en-v1.5 (434M params)
Training method	Task-adaptive pre-training (TAPT) via contrastive fine-tuning
Loss	MultipleNegativesRankingLoss (in-batch negatives)
Training data	FormBench-Triplets — 44,413 (query, anchor, hard-negative) tuples
Embedding dimension	1024
Max sequence length	8192 (training: 2048)
Precision	bf16
Learning rate	1e-5
Per-GPU batch size	64
Epochs	5
Hardware	8× AMD MI250X, DDP

The training-triplet set is reconstructable from the qrel files in Formbench-anon/FormBench following the protocol in §3 of the paper.

Evaluation results

Evaluated on the FormBench test split (n = 5,459 queries) under both corpus variants, following the protocol in §4 of the paper. FAISS exact inner-product search at top-k = 100.

FormBench-Structured (C1) — within-domain near-miss distractors

Metric	Value
Binary nDCG@10	0.3561
MRR (binary qrels)	0.3125
Graded nDCG@10	0.2087
R@100 (binary qrels)	0.7586
FAISS search latency	18.9 ms/query

FormBench-Random (C0) — random-distractor corpus

Metric	Value
Binary nDCG@10	0.4308
MRR (binary qrels)	0.3883
Graded nDCG@10	0.2553
R@100 (binary qrels)	0.8011
FAISS search latency	18.8 ms/query

For reference: BM25 lexical baseline: binary nDCG@10 = 0.3751 (C1), 0.4665 (C0).

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Formbench-anon/gte-large-formbench-mnrl")

passages = [
    "An adhesive composition comprising a styrene-acrylate copolymer ...",
    "A water-based latex paint formulation containing ...",
]
queries = [
    "what wax-seeded latex polymers improve scuff resistance in architectural coatings?",
]

passage_embeds = model.encode(passages, normalize_embeddings=True)
query_embeds   = model.encode(queries,  normalize_embeddings=True)

Intended use

Domain-specific retrieval over formulation patents — adhesives, coatings, lubricants, pharmaceuticals, agrochemicals, personal care, food. Particularly suited to within-domain near-miss discrimination, where general-purpose embedders have been shown to fail.

Limitations

• Training queries are LLM-generated (Sonnet 4 + Haiku 4.5 quality filter) and may not match real practitioner intent.
• Coverage limited to USPTO utility patents (1995–2022) in English only.
• Performance on out-of-domain retrieval is not characterised.

Citation

@misc{formbench2026,
  title  = { {FormBench}: Evaluating Chemical Knowledge Retrieval in Formulation Patents },
  author = { Anonymous Authors },
  year   = { 2026 },
  note   = { Under double-blind review at NeurIPS 2026 Datasets \& Benchmarks Track }
}

License

Apache 2.0, inherited from the base model.