Encoder: microsoft/deberta-v3-xsmall (frozen weights not included — loaded from HuggingFace at runtime) Custom head: biaffine_head.safetensors — word projection, arc/rel/POS MLPs and biaffine layers Labels: 53 DEPREL labels · 19 UPOS tags Score convention: s_arc[dep, head], s_rel[dep, head, rel]

Metrics (EWT + GUM, decode: Eisner (projective MST))

Split	LAS	UPOS	UCM	LCM
dev	93.03%	98.09%	70.75%	60.96%
test	92.97%	98.04%	70.23%	60.91%

ONNX / production use

model.onnx — fp32 model (Recommended for CPU inference). model.fp16.onnx — fp16 model (For GPU or environments with native fp16 support, ~~139 MB).

Inputs: subwords [B, W, 20] int64. Outputs: s_arc [B,W,W], s_rel [B,W,W,R], s_pos [B,W,P].

bash

# download only inference artifacts
hf download ghotriw/deberta-v3-xsmall-biaffine-dep-pos-en-ewt-gum \
  model.fp16.onnx model.onnx vocabs.json tokenizer.json \
  --local-dir ./model

vocabs.json — DEPREL and UPOS vocabularies (str→int dicts).

Input format

The model expects a word-level subword grid [B, W, fix_len=20] (int64), where each word is independently tokenised with the encoder's sentencepiece tokeniser and padded/truncated to 20 subword slots. Position 0 is a synthetic ROOT word whose only subword is [CLS] (id 1).

Vocabularies

config.json contains rel_vocab (str→int) and pos_vocab (str→int). Index 0 is the <pad> / ROOT slot and should be ignored in evaluation.