Enoki OpenIE Encoder

Enoki OpenIE Encoder is an LLM-free fact extractor for English. It distills Enoki's LLM-based OpenIE decomposition into a ModernBERT-large Iterative Grid Labeling encoder that converts sentences into text-anchored (subject, relation, object) triples. Like the atomic-fact decomposition stage in FActScore-style factuality pipelines, it turns free-form text into facts that can be verified independently, but performs the extraction with a single encoder instead of a generative LLM.

The model is designed for Enoki's multi-level hallucination detection pipeline. Extracted facts can be checked against a reference context with a separate NLI or factuality verifier. Because the extracted facts remain anchored to tokens in the original sentence, an unsupported fact can be projected directly back to its source span. The same representation supports both claim-level verification and span-level localization without a separate LLM-based claim-to-text alignment step.

The encoder was trained on the EnokiQA dev split using incremental triples produced by Enoki-LLM, making the training setup a form of fact-extractor distillation.

Source code: s-nlp/Enoki

Paper: Enoki: Efficient Multi-Level Hallucination Detection

Training data: s-nlp/EnokiQA

Installation

pip install torch "transformers>=4.48,<5" nltk

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "s-nlp/enoki-openie-encoder",
    trust_remote_code=True,
)

results = model.extract_triples(
    "Apple acquired Beats Electronics for $3 billion in 2014.",
    min_confidence=0.7,
)
print(results)

Example output:

[
    {
        "sentence": "Apple acquired Beats Electronics for $3 billion in 2014.",
        "triples": [
            {
                "subject": "Apple",
                "relation": "acquired",
                "object": "Beats Electronics",
                "confidence": 0.969,
            },
            {
                "subject": "Apple",
                "relation": "acquired for",
                "object": "$3 billion",
                "confidence": 0.934,
            },
        ],
    }
]

For several sentences, pass a list to model.extract_triples([...]).

Command line

python inference.py \
  --model s-nlp/enoki-openie-encoder \
  --text "Barack Obama was born in Honolulu." \
  --min-confidence 0.7

Hallucination detection with Enoki

Use the Enoki package to combine this encoder with NLI-based fact verification and span-level hallucination localization:

git clone https://github.com/s-nlp/Enoki.git
cd Enoki
poetry install --all-extras

from enoki import EnokiPipeline

context = "Apple acquired Beats Electronics in 2014 for $3 billion."
answer = "Apple acquired Beats Electronics in 2015 for $3 billion."

results = EnokiPipeline(
    model="s-nlp/enoki-openie-encoder"
).detect(context=context, answer=answer)

print(results)

[
    {
        "span": "2015",
        "start": 36,
        "end": 40,
        "fact": {
            "subject": "Apple",
            "predicate": "acquired in",
            "object": "Beats Electronics 2015",
        },
        "probability": 0.9625,
    }
]

Notes

• The model is designed for English text.
• Pass one sentence per input item.
• The default maximum sequence length is 128 tokens.
• IGL can return nested or incremental triples. Use min_confidence=0.7–0.8 when a smaller, higher-precision result set is preferred.
• Loading requires trust_remote_code=True because the IGL architecture and OpenIE decoder are custom Transformers code included in this repository.

Citation

If you use Enoki in your research, please cite:

@misc{rykov2026enokiefficientmultilevelhallucination,
  title         = {Enoki: Efficient Multi-Level Hallucination Detection},
  author        = {Elisei Rykov and Timur Ionov and Nikolay Ivanov and Maksim Savkin and Maksim Makarenko and Alexander Panchenko and Vasily Konovalov and Julia Belikova},
  year          = {2026},
  eprint        = {2609.00581},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.00581},
}