hikmaai-mdeberta-v3-base-xnli-multilingual

hikmaai-mdeberta-v3-base-xnli-multilingual

ONNX repackaging of a multilingual zero-shot NLI model, published for the TOPICALML input control in hikma-mirsad. It scores an input prompt against operator-defined topic labels via zero-shot Natural Language Inference.

Provenance

• Base model: MoritzLaurer/mDeBERTa-v3-base-mnli-xnli (MIT).
• Exported by: scripts/export_xnli_onnx.py in hikma-mirsad (optimum main_export, opset 17; INT8 dynamic quant for CPU, FP16 for GPU).
• License: MIT (inherited from the base model).

How TOPICALML uses it (wire contract)

Zero-shot topic classification is run as NLI: premise = the user prompt, hypothesis = "This text is about <label>.". The gateway takes softmax over the entailment vs contradiction logits (neutral is ignored); the entailment probability is the on-topic score.

The Go scorer (internal/topicalml/scorer.go) depends on this contract, validated at export:

• Inputs: exactly input_ids + attention_mask (mDeBERTa-v3 has type_vocab_size=0, so no token_type_ids).
• Output: 3 logits ordered 0=entailment, 1=neutral, 2=contradiction.

Sanity (premise about weapons): P(entailment | "about weapons") = 0.998, P(entailment | "about cooking") = 0.001.

Layout

Scope

The classifier scores topic presence, not intent: a benign text discussing a blocked topic (e.g. an essay arguing against weapons) scores on-topic. TOPICALML defaults to shadow and requires a calibrated threshold before enforcing.