raphaelmansuy/atlop-roberta-base-onnx

ATLOP RoBERTa-base ONNX (edgextract)

Document-level relation extraction head trained for the edgextract browser demo (Wikidata relations mode). The encoder is roberta-base; the head follows ATLOP (Zhou et al., AAAI 2021): entity markers, localized context pooling, grouped bilinear classifier, adaptive-threshold loss.

Files

Path	Role
model.onnx	Full ATLOP forward (encoder + head), opset 17
config.json	rel2id, margin shift, decision-model band params
rel_info.json	Wikidata P-code → English label / description
tokenizer* / vocab.json / merges.txt	RoBERTa tokenizer (same as training)

ONNX inputs / output

• input_ids [1, seq], attention_mask [1, seq]
• entity_starts [E, M] int64, -1 for unused mention slots (start-marker indices)
• hts [P, 2] int64 entity-pair indices
• output logits [P, 97] (class 0 = threshold / NA, then 96 Wikidata relations)

Training data

Fine-tuned on Re-DocRED train only (train_revised.json). Dev selected the checkpoint and margin shift; test was opened once for the published comparison in the edgextract repo. Do not treat browser extractions as comparable to that score: the benchmark supplies gold entity clusters; the demo finds names from raw text.

How to use in the demo

1. Open the edgextract web demo.
2. Step 2 → Wikidata relations (RoBERTa).
3. Load RoBERTa (default Hub id: raphaelmansuy/atlop-roberta-base-onnx).
4. Extract. Uncertain pairs are verified by the selected decision backend (Tev1 WebGPU or Ollama).

Local export (from the repo):

make demo-atlop-export
make demo-atlop-publish   # this card + weights → Hugging Face

License

• This repository: Apache-2.0
• Base encoder FacebookAI/roberta-base: MIT
• Re-DocRED: see Tan et al., EMNLP 2022 / the dataset card for data terms