---
language: en
library_name: transformers
license: cc-by-nc-2.0
tags:
- scientific-claim-verification
- scifact
- text-classification
- evidence-aware-verification
- retrieval-aware-training
metrics:
- accuracy
- f1
---

# VeriSci Claim Verifier Dense-Retrieval-Adapted Candidate

Dense-retrieval-adapted VeriSci classifier for scientific
claim verification over `SUPPORTS`, `REFUTES`, and `NOT_ENOUGH_INFO`.

## Promotion Status

Not promoted. This checkpoint was trained and evaluated successfully, but it
regressed against `rishhh/verisci-claim-verifier-retrieval-adapted-seed123` on
both the derived test split and the dense/hybrid retrieval-grounded SciFact
validation gate.

## Training Data

- `allenai/scifact_entailment` train split with gold evidence sentences.
- `allenai/scifact` train claims converted into gold evidence snippets,
  BM25-selected evidence snippets, and BM25-retrieved hard `NOT_ENOUGH_INFO`
  negatives.
- `andreiaalexa/scifact-relevance-pairs` title/train hard negatives.

## Derived Test

Accuracy: 0.8935
Macro F1: 0.8567

## Full SciFact Validation Gate

Top-1 end-to-end accuracy: 0.5711
Top-1 end-to-end macro F1: 0.5536
Top-5 decisive accuracy: 0.6356
Top-5 decisive macro F1: 0.5935

## Dense/Hybrid Retrieval-Grounded Gate

Using `rishhh/verisci-scifact-e5-retriever`, hybrid alpha `0.75`, and guarded
top-5 decisive aggregation:

- Accuracy: 0.6756
- Macro F1: 0.6426

The current promoted retrieval-adapted verifier reaches 0.6911 accuracy and
0.6529 macro F1 under the same dense/hybrid guarded gate.

See `evaluation/eval_summary.json` for full metrics and limitations.

## Responsible Use

This model is not a medical device, not a substitute for peer review, and should
not be used for clinical, legal, or public-policy decisions without expert
review.