Base model: XLM-RoBERTa Base (multilingual, 10+ languages)
Training data: FewNERD fine-grained supervised
Entity types: 66 types + O (including building-hotel, building-airport, location-GPE, location-park, location-bodiesofwater, location-mountain, etc.)
Quantization: INT8 dynamic quantization on encoder

Files

File	Size	Description
`encoder.onnx`	~266 MB	XLM-RoBERTa encoder (INT8 quantized)
`classifier.onnx`	0.4 MB	Linear classifier (1536 -> 67 classes)
`labels.json`	2 KB	ID to label mapping
`tokenizer.json`	17 MB	SentencePiece tokenizer

Inference

SpanMarker uses a custom architecture:

Tokenize input and enumerate all spans (up to entity_max_length=8 tokens)
Insert start/end marker tokens into the sequence
Run encoder to get hidden states
Extract start marker embeddings from [start_idx : start_idx + n_pairs]
Extract end marker embeddings from [start_idx + n_pairs : start_idx + n_pairs*2]
Concatenate start+end -> (n_pairs, 1536) features
Pad to (256, 1536) and run classifier -> (256, 67) logits
Softmax + filter non-O predictions above threshold

Fixed sequence length of 512 (batch=1)
Longer texts need to be chunked (the original model does this internally with batching)
INT8 quantization slightly reduces scores (~0.05-0.1 lower) but maintains correct predictions

CC-BY-SA-4.0 (same as original model)