v1.1: multi-sentence claim data (see card: Versions)

97612bd verified about 1 month ago

8.84 kB

---
license: apache-2.0
language:
- multilingual
- en
- de
- fr
- es
- pt
- it
- nl
- pl
- ru
- uk
- tr
- ar
- hi
- zh
- ja
- ko
- vi
- id
- th
library_name: transformers
pipeline_tag: text-classification
base_model: jhu-clsp/mmBERT-base
tags:
- hallucination-detection
- groundedness
- faithfulness
- fact-checking
- rag
- llm-evaluation
- guardrails
- nli
- modernbert
- mmbert
- onnx
- transformers.js
datasets:
- wandb/RAGTruth-processed
- alisawuffles/WANLI
---
# Hallucination Guard (base, 308M)

A small, **multilingual groundedness checker**. Given a source (retrieved documents, a transcript, a tool result) and
an AI response or claim, it predicts whether the response is **supported** by the source. Use it to flag RAG answers,
summaries and agent outputs that add, change or contradict facts.

- **Multilingual**: trained on data in 30 languages. On HaluEval translated into six languages it stays at the level
  of its English score, while English-only checkers drop (table below).
- **Long context**: 8k-token window (trained at 2k); long sources can also be chunked (see below).
- **Open**: Apache-2.0, ungated. Trained on permissively licensed data and Qwen-generated synthetic data. ONNX
  included.
- **Honest about where it loses**: on English claim-level fact-checking (LLM-AggreFact), MiniCheck and HHEM are
  better. If you only need English, compare them on your data.

Try it in the browser: [Horizon-Labs/hallucination-guard demo](https://huggingface.co/spaces/Horizon-Labs/hallucination-guard).

Part of [Agent I/O Guards](https://huggingface.co/collections/Horizon-Labs/agent-i-o-guards-6ab403c49494bc2b71ca7669)
(prompt injection, PII, groundedness). Source code: [github.com/horizon-ai-labs/agent-io-guards](https://github.com/horizon-ai-labs/agent-io-guards).

## Quick start

```python
from transformers import pipeline
clf = pipeline("text-classification", model="Horizon-Labs/hallucination-guard-base")
doc = "The Eiffel Tower is 330 metres tall and was completed in 1889 for the World's Fair."
clf({"text": doc, "text_pair": "The tower was finished in 1889."})   # SUPPORTED
clf({"text": doc, "text_pair": "The tower was finished in 1899."})   # UNSUPPORTED
```

Put the **source first** and the **response or claim second**. Labels: `UNSUPPORTED` (0) and `SUPPORTED` (1).
The score for `SUPPORTED` is the support probability.

For long sources, or to find *which* sentence is unsupported, split the response into sentences and score each one
against the source. For a response-level score, take the minimum over sentences:

```python
import re
def check(source, response, clf=clf):
    sents = [s for s in re.split(r"(?<=[.!?。！？])\s+", response) if s.strip()]
    res = clf([{"text": source, "text_pair": s} for s in sents], truncation=True, max_length=2048)
    return [(s, r["label"], r["score"]) for s, r in zip(sents, res)]
```

## Evaluation

Balanced accuracy at a 0.5 threshold. All models were run by us with the same script
(`ground/evaluate_ground.py`). MiniCheck used context chunking with max over chunks, as its own library does. HHEM
used contexts capped at 12,000 characters, because it ran out of memory on the longest documents. † marks sets that are
**in-distribution for our models**: we trained on RAGTruth's train split, and those rows use its test split.

### LLM-AggreFact (English claim verification, 11 datasets, up to 1,000 examples each)

| | **this model** | small (141M) | MiniCheck-RoBERTa-L | MiniCheck-DeBERTa-L | HHEM-2.1-open |
|---|---|---|---|---|---|
| AggreFact-CNN | 0.618 | 0.607 | **0.664** | 0.622 | 0.629 |
| AggreFact-XSum | 0.622 | 0.644 | **0.709** | 0.688 | 0.706 |
| ClaimVerify | 0.737 | 0.682 | **0.783** | 0.744 | 0.753 |
| ExpertQA | 0.593 | 0.553 | **0.606** | 0.597 | 0.568 |
| FactCheck-GPT | 0.674 | 0.657 | **0.754** | 0.728 | 0.729 |
| LFQA | 0.813 | 0.726 | **0.863** | 0.838 | 0.839 |
| RAGTruth † | **0.815** | 0.796 | 0.788 | 0.773 | 0.734 |
| Reveal | 0.834 | 0.798 | **0.896** | 0.871 | 0.865 |
| TofuEval-MediaS | **0.707** | 0.654 | 0.696 | 0.676 | 0.679 |
| TofuEval-MeetB | 0.736 | 0.677 | **0.765** | 0.724 | 0.721 |
| Wice | 0.732 | 0.719 | 0.717 | 0.664 | **0.751** |
| **Mean** | 0.716 | 0.683 | **0.749** | 0.721 | 0.725 |

### Multilingual: HaluEval QA and dialogue, translated (400 items per language)

The items were machine-translated with Qwen3.8-27B, keeping their labels. They are disjoint from the English HaluEval
items above. Caveat: the translations come from the same model family we used to generate synthetic training data,
which may favour our model somewhat.

| | **this model** | small (141M) | MiniCheck-RoBERTa-L | MiniCheck-DeBERTa-L | HHEM-2.1-open |
|---|---|---|---|---|---|
| English | **0.720** | 0.657 | 0.589 | 0.670 | 0.701 |
| German | **0.724** | 0.719 | 0.691 | 0.672 | 0.670 |
| Spanish | **0.714** | 0.691 | 0.650 | 0.675 | 0.675 |
| Chinese | **0.738** | 0.713 | 0.627 | 0.621 | 0.546 |
| Japanese | **0.751** | 0.729 | 0.626 | 0.669 | 0.546 |
| Arabic | **0.710** | 0.700 | 0.616 | 0.691 | 0.544 |
| Hindi | **0.740** | 0.730 | 0.595 | 0.670 | 0.564 |
| **Mean of the 6 non-English languages** | **0.729** | 0.714 | 0.634 | 0.666 | 0.591 |

### Versions

v1.1 (this version) adds about 147k multi-sentence claim pairs (claims whose facts are spread over several sentences of
the source, and the same pairs with one needed sentence removed). It is better on English claim-level fact-checking
(LLM-AggreFact). On translated HaluEval it scores lower than v1.0, but a large part of that gap is training noise:
retraining the v1.0 recipe with another random seed gives very different HaluEval numbers (rows below). The translated-
HaluEval AUC is consistently about 0.02 lower with the v1.1 data, so part of the difference is real. Pin the previous
model with `revision="v1.0"`.

| | AggreFact bacc | AggreFact AUC | 6 languages bacc | 6 languages AUC | HaluEval QA | RAGTruth |
|---|---|---|---|---|---|---|
| v1.0 (released) | 0.697 | 0.773 | 0.755 | 0.802 | 0.817 | 0.821 |
| v1.0 recipe, another seed | 0.702 | 0.768 | 0.684 | 0.786 | 0.738 | 0.809 |
| **v1.1 (this version)** | 0.716 | 0.796 | 0.729 | 0.774 | 0.782 | 0.829 |
| v1.1 recipe, another seed | 0.711 | 0.794 | 0.706 | 0.777 | 0.714 | 0.822 |

### Other benchmarks

| | **this model** | small (141M) | MiniCheck-RoBERTa-L | MiniCheck-DeBERTa-L | HHEM-2.1-open |
|---|---|---|---|---|---|
| HaluEval QA | **0.782** | 0.691 | 0.635 | 0.768 | 0.762 |
| HaluEval dialogue | 0.635 | **0.643** | 0.522 | 0.524 | 0.625 |
| HaluEval summarization | 0.559 | 0.548 | **0.648** | 0.623 | 0.526 |
| RAGTruth test, response level † | **0.829** | 0.798 | 0.604 | 0.626 | 0.750 |

## Limitations

- On English claim-level fact-checking, MiniCheck-RoBERTa-L (0.749) and HHEM (0.725) beat this model
  (0.716) on LLM-AggreFact. Our advantage is in other languages, and in QA and dialogue grounding.
- It judges support by the given source only. It is not a world-knowledge fact checker: a true statement that the
  source doesn't contain is `UNSUPPORTED`.
- Simple arithmetic or temporal inferences are often marked `UNSUPPORTED`. For example, "opened before 2022" given a
  source that says "opened in March 2021". The small model also misses some paraphrases ("weekdays" for "Monday to
  Friday") that the base model handles.
- Summaries with many small details (HaluEval summarization) and expert long-form answers (ExpertQA) are hard for
  every model here.
- Much of the training data is synthetic (Qwen3.8-27B). The multilingual numbers come from translated data, not native
  benchmarks.

## Training

- Backbone: [jhu-clsp/mmBERT-base](https://huggingface.co/jhu-clsp/mmBERT-base) (MIT), sequence-pair classification, max length 2048,
  bf16.
- About 340k (source, response) pairs:
  - Qwen3.8-27B-generated responses over FineWeb-Edu / FineWeb-2 passages (ODC-BY) in 29 languages: supported
    answers and minimally edited unsupported variants, at both response and sentence level.
  - Qwen-generated documents in 18 genres (news, meeting transcripts, support chats, retrieved snippets, reviews,
    contracts…) with supported and unsupported claims, in 30 languages.
  - [RAGTruth](https://huggingface.co/datasets/wandb/RAGTruth-processed) train split (MIT), response level.
  - [WANLI](https://huggingface.co/datasets/alisawuffles/WANLI) (CC-BY-4.0).
  - (v1.1) About 147k multi-sentence claim pairs generated with Qwen3.8-27B (`code/ground/gen_c2d.py`): invented documents whose facts are spread over different sentences, and FineWeb / FineWeb-2 passages with claims that combine 2-3 sentences; unsupported versions remove one needed sentence. About 40% English.
- Not used: ANLI and other non-commercial NLI data, DocNLI (derived from non-commercial sources), and every benchmark
  above except RAGTruth's train split.

	---
	license: apache-2.0
	language:
	- multilingual
	- en
	- de
	- fr
	- es
	- pt
	- it
	- nl
	- pl
	- ru
	- uk
	- tr
	- ar
	- hi
	- zh
	- ja
	- ko
	- vi
	- id
	- th
	library_name: transformers
	pipeline_tag: text-classification
	base_model: jhu-clsp/mmBERT-base
	tags:
	- hallucination-detection
	- groundedness
	- faithfulness
	- fact-checking
	- rag
	- llm-evaluation
	- guardrails
	- nli
	- modernbert
	- mmbert
	- onnx
	- transformers.js
	datasets:
	- wandb/RAGTruth-processed
	- alisawuffles/WANLI
	---

	# Hallucination Guard (base, 308M)

	A small, multilingual groundedness checker. Given a source (retrieved documents, a transcript, a tool result) and
	an AI response or claim, it predicts whether the response is supported by the source. Use it to flag RAG answers,
	summaries and agent outputs that add, change or contradict facts.

	- Multilingual: trained on data in 30 languages. On HaluEval translated into six languages it stays at the level
	of its English score, while English-only checkers drop (table below).
	- Long context: 8k-token window (trained at 2k); long sources can also be chunked (see below).
	- Open: Apache-2.0, ungated. Trained on permissively licensed data and Qwen-generated synthetic data. ONNX
	included.
	- Honest about where it loses: on English claim-level fact-checking (LLM-AggreFact), MiniCheck and HHEM are
	better. If you only need English, compare them on your data.

	Try it in the browser: [Horizon-Labs/hallucination-guard demo](https://huggingface.co/spaces/Horizon-Labs/hallucination-guard).

	Part of [Agent I/O Guards](https://huggingface.co/collections/Horizon-Labs/agent-i-o-guards-6ab403c49494bc2b71ca7669)
	(prompt injection, PII, groundedness). Source code: [github.com/horizon-ai-labs/agent-io-guards](https://github.com/horizon-ai-labs/agent-io-guards).

	## Quick start

	```python
	from transformers import pipeline

	clf = pipeline("text-classification", model="Horizon-Labs/hallucination-guard-base")
	doc = "The Eiffel Tower is 330 metres tall and was completed in 1889 for the World's Fair."
	clf({"text": doc, "text_pair": "The tower was finished in 1889."}) # SUPPORTED
	clf({"text": doc, "text_pair": "The tower was finished in 1899."}) # UNSUPPORTED
	```

	Put the source first and the response or claim second. Labels: `UNSUPPORTED` (0) and `SUPPORTED` (1).
	The score for `SUPPORTED` is the support probability.

	For long sources, or to find which sentence is unsupported, split the response into sentences and score each one
	against the source. For a response-level score, take the minimum over sentences:

	```python
	import re
	def check(source, response, clf=clf):
	sents = [s for s in re.split(r"(?<=[.!?。！？])\s+", response) if s.strip()]
	res = clf([{"text": source, "text_pair": s} for s in sents], truncation=True, max_length=2048)
	return [(s, r["label"], r["score"]) for s, r in zip(sents, res)]
	```

	## Evaluation

	Balanced accuracy at a 0.5 threshold. All models were run by us with the same script
	(`ground/evaluate_ground.py`). MiniCheck used context chunking with max over chunks, as its own library does. HHEM
	used contexts capped at 12,000 characters, because it ran out of memory on the longest documents. † marks sets that are
	in-distribution for our models: we trained on RAGTruth's train split, and those rows use its test split.

	### LLM-AggreFact (English claim verification, 11 datasets, up to 1,000 examples each)

	\| \| this model \| small (141M) \| MiniCheck-RoBERTa-L \| MiniCheck-DeBERTa-L \| HHEM-2.1-open \|
	\|---\|---\|---\|---\|---\|---\|
	\| AggreFact-CNN \| 0.618 \| 0.607 \| 0.664 \| 0.622 \| 0.629 \|
	\| AggreFact-XSum \| 0.622 \| 0.644 \| 0.709 \| 0.688 \| 0.706 \|
	\| ClaimVerify \| 0.737 \| 0.682 \| 0.783 \| 0.744 \| 0.753 \|
	\| ExpertQA \| 0.593 \| 0.553 \| 0.606 \| 0.597 \| 0.568 \|
	\| FactCheck-GPT \| 0.674 \| 0.657 \| 0.754 \| 0.728 \| 0.729 \|
	\| LFQA \| 0.813 \| 0.726 \| 0.863 \| 0.838 \| 0.839 \|
	\| RAGTruth † \| 0.815 \| 0.796 \| 0.788 \| 0.773 \| 0.734 \|
	\| Reveal \| 0.834 \| 0.798 \| 0.896 \| 0.871 \| 0.865 \|
	\| TofuEval-MediaS \| 0.707 \| 0.654 \| 0.696 \| 0.676 \| 0.679 \|
	\| TofuEval-MeetB \| 0.736 \| 0.677 \| 0.765 \| 0.724 \| 0.721 \|
	\| Wice \| 0.732 \| 0.719 \| 0.717 \| 0.664 \| 0.751 \|
	\| Mean \| 0.716 \| 0.683 \| 0.749 \| 0.721 \| 0.725 \|

	### Multilingual: HaluEval QA and dialogue, translated (400 items per language)

	The items were machine-translated with Qwen3.8-27B, keeping their labels. They are disjoint from the English HaluEval
	items above. Caveat: the translations come from the same model family we used to generate synthetic training data,
	which may favour our model somewhat.

	\| \| this model \| small (141M) \| MiniCheck-RoBERTa-L \| MiniCheck-DeBERTa-L \| HHEM-2.1-open \|
	\|---\|---\|---\|---\|---\|---\|
	\| English \| 0.720 \| 0.657 \| 0.589 \| 0.670 \| 0.701 \|
	\| German \| 0.724 \| 0.719 \| 0.691 \| 0.672 \| 0.670 \|
	\| Spanish \| 0.714 \| 0.691 \| 0.650 \| 0.675 \| 0.675 \|
	\| Chinese \| 0.738 \| 0.713 \| 0.627 \| 0.621 \| 0.546 \|
	\| Japanese \| 0.751 \| 0.729 \| 0.626 \| 0.669 \| 0.546 \|
	\| Arabic \| 0.710 \| 0.700 \| 0.616 \| 0.691 \| 0.544 \|
	\| Hindi \| 0.740 \| 0.730 \| 0.595 \| 0.670 \| 0.564 \|
	\| Mean of the 6 non-English languages \| 0.729 \| 0.714 \| 0.634 \| 0.666 \| 0.591 \|

	### Versions

	v1.1 (this version) adds about 147k multi-sentence claim pairs (claims whose facts are spread over several sentences of
	the source, and the same pairs with one needed sentence removed). It is better on English claim-level fact-checking
	(LLM-AggreFact). On translated HaluEval it scores lower than v1.0, but a large part of that gap is training noise:
	retraining the v1.0 recipe with another random seed gives very different HaluEval numbers (rows below). The translated-
	HaluEval AUC is consistently about 0.02 lower with the v1.1 data, so part of the difference is real. Pin the previous
	model with `revision="v1.0"`.

	\| \| AggreFact bacc \| AggreFact AUC \| 6 languages bacc \| 6 languages AUC \| HaluEval QA \| RAGTruth \|
	\|---\|---\|---\|---\|---\|---\|---\|
	\| v1.0 (released) \| 0.697 \| 0.773 \| 0.755 \| 0.802 \| 0.817 \| 0.821 \|
	\| v1.0 recipe, another seed \| 0.702 \| 0.768 \| 0.684 \| 0.786 \| 0.738 \| 0.809 \|
	\| v1.1 (this version) \| 0.716 \| 0.796 \| 0.729 \| 0.774 \| 0.782 \| 0.829 \|
	\| v1.1 recipe, another seed \| 0.711 \| 0.794 \| 0.706 \| 0.777 \| 0.714 \| 0.822 \|

	### Other benchmarks

	\| \| this model \| small (141M) \| MiniCheck-RoBERTa-L \| MiniCheck-DeBERTa-L \| HHEM-2.1-open \|
	\|---\|---\|---\|---\|---\|---\|
	\| HaluEval QA \| 0.782 \| 0.691 \| 0.635 \| 0.768 \| 0.762 \|
	\| HaluEval dialogue \| 0.635 \| 0.643 \| 0.522 \| 0.524 \| 0.625 \|
	\| HaluEval summarization \| 0.559 \| 0.548 \| 0.648 \| 0.623 \| 0.526 \|
	\| RAGTruth test, response level † \| 0.829 \| 0.798 \| 0.604 \| 0.626 \| 0.750 \|

	## Limitations

	- On English claim-level fact-checking, MiniCheck-RoBERTa-L (0.749) and HHEM (0.725) beat this model
	(0.716) on LLM-AggreFact. Our advantage is in other languages, and in QA and dialogue grounding.
	- It judges support by the given source only. It is not a world-knowledge fact checker: a true statement that the
	source doesn't contain is `UNSUPPORTED`.
	- Simple arithmetic or temporal inferences are often marked `UNSUPPORTED`. For example, "opened before 2022" given a
	source that says "opened in March 2021". The small model also misses some paraphrases ("weekdays" for "Monday to
	Friday") that the base model handles.
	- Summaries with many small details (HaluEval summarization) and expert long-form answers (ExpertQA) are hard for
	every model here.
	- Much of the training data is synthetic (Qwen3.8-27B). The multilingual numbers come from translated data, not native
	benchmarks.

	## Training

	- Backbone: [jhu-clsp/mmBERT-base](https://huggingface.co/jhu-clsp/mmBERT-base) (MIT), sequence-pair classification, max length 2048,
	bf16.
	- About 340k (source, response) pairs:
	- Qwen3.8-27B-generated responses over FineWeb-Edu / FineWeb-2 passages (ODC-BY) in 29 languages: supported
	answers and minimally edited unsupported variants, at both response and sentence level.
	- Qwen-generated documents in 18 genres (news, meeting transcripts, support chats, retrieved snippets, reviews,
	contracts…) with supported and unsupported claims, in 30 languages.
	- [RAGTruth](https://huggingface.co/datasets/wandb/RAGTruth-processed) train split (MIT), response level.
	- [WANLI](https://huggingface.co/datasets/alisawuffles/WANLI) (CC-BY-4.0).
	- (v1.1) About 147k multi-sentence claim pairs generated with Qwen3.8-27B (`code/ground/gen_c2d.py`): invented documents whose facts are spread over different sentences, and FineWeb / FineWeb-2 passages with claims that combine 2-3 sentences; unsupported versions remove one needed sentence. About 40% English.
	- Not used: ANLI and other non-commercial NLI data, DocNLI (derived from non-commercial sources), and every benchmark
	above except RAGTruth's train split.