nicpopovic/jedi-deberta-v3-large

JEDI (DeBERTa-v3-large)

Checkpoint of JEDI from the EMNLP 2025 paper Extractive Fact Decomposition for Interpretable Natural Language Inference in one Forward Pass.

• Code: https://github.com/nicpopovic/JEDI
• Demo: https://huggingface.co/spaces/nicpopovic/JEDI
• Backbone: microsoft/deberta-v3-large
• more information: https://jedi.nicpopovic.com
• arxiv: https://arxiv.org/abs/2509.18901
• aclanthology: https://aclanthology.org/2025.emnlp-main.1615/

weights.ckpt is a PyTorch state dict for IENLIModel (jedi/modeling.py), which wraps AutoModelForTokenClassification of the backbone. It cannot be loaded with AutoModel.from_pretrained.

from huggingface_hub import hf_hub_download
path = hf_hub_download("nicpopovic/jedi-deberta-v3-large", "weights.ckpt")

python -m jedi.evaluate --task anli_r3 --model microsoft/deberta-v3-large --checkpoints <path>

Results

Accuracy (%) of this checkpoint (single run).

	ANLI R1	R2	R3	ANLI	ConTRoL	RTE	WNLI	HANS
this checkpoint	76.9	63.5	58.7	65.9	53.6	88.5	77.5	76.9

Citation

@inproceedings{popovic-farber-2025-extractive,
    title = "Extractive Fact Decomposition for Interpretable Natural Language Inference in one Forward Pass",
    author = {Popovi{\v{c}}, Nicholas  and
      F{\"a}rber, Michael},
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1615/",
    doi = "10.18653/v1/2025.emnlp-main.1615",
    pages = "31692--31705",
    ISBN = "979-8-89176-332-6",
    abstract = "Recent works in Natural Language Inference (NLI) and related tasks, such as automated fact-checking, employ atomic fact decomposition to enhance interpretability and robustness. For this, existing methods rely on resource-intensive generative large language models (LLMs) to perform decomposition. We propose JEDI, an encoder-only architecture that jointly performs extractive atomic fact decomposition and interpretable inference without requiring generative models during inference. To facilitate training, we produce a large corpus of synthetic rationales covering multiple NLI benchmarks. Experimental results demonstrate that JEDI achieves competitive accuracy in distribution and significantly improves robustness out of distribution and in adversarial settings over models based solely on extractive rationale supervision. Our findings show that interpretability and robust generalization in NLI can be realized using encoder-only architectures and synthetic rationales."
}