grounding-en
Does the document actually support this claim? grounding-en is a cross-encoder that scores whether
a hypothesis (a number, date, or fact) is entailed by a premise drawn from a real document — a
financial table, a filing, prose evidence.
It is the open, English member of Nutrient's grounding model family.
- 🎯 Try it: grounding-demo
- 🏆 Leaderboard: grounding-leaderboard
- 📊 Benchmark: grounding-benchmark
Results
On the held-out English grounding-benchmark (ROC-AUC), against the strongest open English NLI models:
| Facet | grounding-en | grounding-multilingual | DeBERTa-v3-large zero-shot | DeBERTa-v3-large MNLI/FEVER/ANLI | BART-large MNLI |
|---|---|---|---|---|---|
| Overall | .882 | .925 | .786 | .769 | .636 |
| Number | .923 | .969 | .658 | .642 | .478 |
| Date | .998 | .999 | .995 | .995 | .924 |
| String | .955 | .949 | .941 | .913 | .757 |
| Table premises | .863 | .915 | .766 | .747 | .611 |
| Prose premises | .964 | .970 | .929 | .945 | .892 |
grounding-en leads the field on the hard axis — number grounding .92 vs .48–.66 for general-purpose
NLI models — while matching or beating them everywhere else. Full ranking on the
leaderboard. The commercial sibling
grounding-multilingual scores a touch
higher again and covers 15+ languages.
Usage
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
m = "nutrientdocs/grounding-en"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSequenceClassification.from_pretrained(m).eval()
premise = "Revenue | 2023 | $4,213M\nRevenue | 2022 | $3,905M"
hypothesis = "2023 revenue was $4.2 billion."
enc = tok(premise, hypothesis, truncation=True, max_length=1024, return_tensors="pt")
with torch.no_grad():
probs = torch.softmax(model(**enc).logits, dim=-1)[0]
p_support = probs[0].item() # entailment is class index 0 (id2label = {0: entailment, 1: not_entailment})
print(f"grounded support = {p_support:.3f}")
An ONNX export is provided under onnx/ for on-device / ONNX Runtime deployment.
Calibrating the score
Fine-tuning maximizes ranking (AUC), which tends to make the raw probability overconfident. For a
score you can gate on ("0.9 means ~90% right"), apply temperature scaling — divide the logits by a
fitted T before softmax. It's monotonic, so it leaves AUC/ranking untouched and only repairs the
confidence values. On the serving distribution we fit T = 1.29 (ECE 0.028 → 0.009). Re-fit T on
your distribution whenever your input pipeline changes.
Intended use & limits
- Use it for: verifying extracted values against source documents, hallucination/citation checking, routing low-confidence extractions for review.
- Limits: English only. The remaining ceiling is reasoning table-claim negatives and multi-step arithmetic. As a dedicated grounding model it trades a little general-NLI accuracy for grounding.
License & training data
Weights are Apache-2.0. Trained on a multi-corpus grounding set. The public, redistributable slice of the evaluation data is grounding-benchmark (CC-BY-SA-4.0); the full training set is not redistributed.
About the author
This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.