nli-mdeberta-en — English-pruned NLI cross-encoder for Statelet

English-pruned ONNX build of MoritzLaurer/mDeBERTa-v3-base-mnli-xnli, packaged for the NLI discrimination paths in Statelet: conflict detection, query–fact relevance, and answer-chain composition.

This is an English-only derivative. Do not use it for multilingual NLI.

⚠️ Licensing — read before commercial use

The upstream weights are published under MIT, and that is the license declared here. However the upstream model is fine-tuned on MultiNLI and XNLI, and:

• XNLI is released under CC BY-NC 4.0 — a non-commercial license (see LICENSE in facebookresearch/XNLI).
• MultiNLI is mixed: CC-BY-3.0 / CC-BY-SA-3.0 / MIT / other.

Whether a non-commercial training-data term restricts the resulting weights is unsettled, and the upstream author chose MIT. We surface this rather than resolve it. Review it yourself before commercial deployment.

Label order

config.json here carries only the label mapping, which is identical to upstream:

0 = entailment, 1 = neutral, 2 = contradiction

Statelet reads id2label from config.json; this ordering is not the same as some other NLI checkpoints, so do not assume it when swapping models.

What was changed

The checkpoint is fully fp32 — fp16 breaks it via ONNX Runtime's SimplifiedLayerNormFusion, so it was never fp16'd. About 770 MB of it is the 250,000-row multilingual word-embedding table (deberta.embeddings.word_embeddings.weight), consumed by a single Gather.

That table is pruned to its ~120k Latin / shared / byte / special rows, and the kept rows are stored INT8. The INT8 surgery swaps only the embedding Gather (adding Gather(int8) + Cast + Mul) and leaves every LayerNorm in fp32, so it does not trigger the fp16 LayerNorm bug. Matmuls stay fp32.

Verification: the pruned tokenizer's ids equal remap(original ids) on English text, and new_table[remap(id)] == old_table[id] by construction, so the fp32 prune is provably lossless for English. The 3-class decision and probabilities were additionally compared fp32 vs pruned-INT8.

Files

file	size	notes
model.onnx	~432 MiB	fp32 embedding table
model_int8.onnx	~242 MiB	INT8 embedding table, fp32 elsewhere
tokenizer.json	—	pruned tokenizer — must be used with these graphs
config.json	—	id2label / label2id only

Use in Statelet

STATELET_MODEL_ROOT=/absolute/path/to/models

with this repository at $STATELET_MODEL_ROOT/nli-mdeberta-en, or set STATELET_NLI_MODEL to point straight at it. If no NLI model is found, Statelet disables NLI and falls back to heuristic conflict/relevance scoring.

Citation

Cite the upstream model and the MNLI/XNLI datasets; see the upstream card.