---
license: other
license_name: nutrient-commercial
library_name: transformers
pipeline_tag: text-classification
language:
  - multilingual
  - en
  - de
  - fr
  - es
  - zh
  - ja
  - ar
  - hi
  - ru
  - tr
  - vi
  - ko
  - sw
  - ur
  - th
tags:
  - grounding
  - hallucination-detection
  - fact-verification
  - nli
  - zero-shot-classification
  - multilingual
  - document-ai
  - cross-encoder
datasets:
  - nutrientdocs/grounding-benchmark
metrics:
  - roc_auc
---

# grounding-multilingual · _commercial_

**Does the document actually support this claim — in 15+ languages?** `grounding-multilingual` 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 strongest, multilingual member of Nutrient's grounding model family. **Weights are commercial**
(not downloadable here); this page is a spec + scorecard. For the open, English model see
[`grounding-en`](https://huggingface.co/nutrientdocs/grounding-en).

- 🎯 **Try it:** [grounding-demo](https://huggingface.co/spaces/nutrientdocs/grounding-demo?model=multi)
- 🏆 **Leaderboard:** [grounding-leaderboard](https://huggingface.co/spaces/nutrientdocs/grounding-leaderboard)
- 📊 **Benchmark:** [grounding-benchmark](https://huggingface.co/datasets/nutrientdocs/grounding-benchmark)

## Results

On the held-out multilingual [grounding-benchmark](https://huggingface.co/datasets/nutrientdocs/grounding-benchmark)
(ROC-AUC), against the top open multilingual NLI models:

| Facet | `grounding-multilingual` | mDeBERTa-v3-base XNLI-2mil7 | mDeBERTa-v3-base MNLI-XNLI | XLM-R-large XNLI |
| --- | ---: | ---: | ---: | ---: |
| **Overall** | **.965** | .894 | .831 | .811 |
| Number | **.999** | .775 | .806 | .765 |
| Table premises | **.999** | .727 | .780 | .715 |
| Prose premises | .926 | **.961** | .864 | .869 |

It leads on **number** (.999 vs .77–.81) and **table** grounding (.999 vs .72–.78) across 15+ languages;
on general prose NLI the best multilingual zero-shot model edges it. Full ranking on the
[leaderboard](https://huggingface.co/spaces/nutrientdocs/grounding-leaderboard). (Date/string grounding
is too rare in the multilingual corpus to score reliably, so it's omitted here.)

## Calibrating the score

Fine-tuning maximizes _ranking_ (AUC), which can leave 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 AUC/ranking is untouched and only the confidence values are repaired.
On the serving distribution we fit **T = 0.94** (ECE 0.012 → 0.007). 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,
  and routing low-confidence extractions for review — across 15+ languages, on-prem.
- **Limits:** the remaining ceiling is _reasoning_ table-claim negatives and multi-step
  arithmetic. As a dedicated grounding checkpoint it trades a little general-NLI accuracy for grounding.

## License & data

The model **weights** are offered under a commercial Nutrient license — on-prem, so documents never leave
your infrastructure. The training set is not redistributed. The **evaluation** data is public and
reproducible: [grounding-benchmark](https://huggingface.co/datasets/nutrientdocs/grounding-benchmark)
(CC-BY-SA-4.0); the full training set is not redistributed.

> ### 📩 Get access
>
> `grounding-multilingual` is commercial and its weights are not downloadable here. To run it on-prem —
> multilingual, calibrated, private — **contact Nutrient: [nutrient.io/contact-sales](https://www.nutrient.io/contact-sales/).**

## About the author

<a href="https://nutrient.io/">
  <img src="https://avatars2.githubusercontent.com/u/1527679?v=3&s=200" height="80" />
</a>

This project is maintained and funded by [Nutrient](https://nutrient.io/) - 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.