arkilite/xlm-roberta-large-xnli-coreml

xlm-roberta-large-xnli — Core ML

Core ML conversion of joeddav/xlm-roberta-large-xnli, a multilingual 3-class Natural Language Inference model (contradiction / neutral / entailment), built on FacebookAI/xlm-roberta-large and fine-tuned on MultiNLI + XNLI.

This repository only re-packages the original model for on-device inference on Apple platforms (Core ML). The weights and all credit belong to the original authors — see Attribution & license below.

Contents

Path	Description
xlm-roberta-large-xnli.mlpackage/	Core ML model bundle (.mlpackage)
tokenizer/	SentencePiece tokenizer: tokenizer.json, sentencepiece.bpe.model, tokenizer_config.json, special_tokens_map.json

Conversion notes

• Format: Core ML (.mlpackage); runs on the Apple Neural Engine / GPU / CPU through the Core ML runtime.
• Fixed-shape input of 512 tokens: the model expects input_ids and attention_mask of length exactly 512 — pad or truncate accordingly.
• Output: 3 logits in the order contradiction (0), neutral (1), entailment (2) — apply softmax for probabilities. (Same id2label as the original model.)
• Format conversion only: no re-training, no change to the weights.

Attribution & license

• Original model: joeddav/xlm-roberta-large-xnli by Joe Davison — license MIT.
• Base model: FacebookAI/xlm-roberta-large — MIT.
• Training data: MultiNLI and XNLI.
• Paper: Unsupervised Cross-lingual Representation Learning at Scale (XLM-R).

This Core ML conversion is distributed under the same MIT license as the original model. If you use it, please credit the original authors above.