rajarshi-saha-2314/vox-parse-distilbert

VoxParse -- Joint Intent + Slot-Filling (DistilBERT)

Fine-tuned distilbert-base-uncased with two task heads -- intent classification on [CLS] and BIO slot tagging on every token -- trained jointly on the SNIPS NLU benchmark.

Part of the VoxParse project: a resume project demonstrating a real-time voice-assistant NLU pipeline. See the project README for full architecture details, training setup, and how to run inference.

• Live demo: https://vox-parse.streamlit.app/
• Code: https://github.com/rajarshi-saha-2314/vox-parse

Test set metrics

task	metric	value
Intent classification	accuracy	97.71%
Intent classification	macro F1	97.77%
Slot filling (entity-level, seqeval)	precision	92.83%
Slot filling (entity-level, seqeval)	recall	94.69%
Slot filling (entity-level, seqeval)	F1	93.75%

Files

• model.pt -- PyTorch state dict for the joint model (encoder + intent head + slot head)
• config.json -- architecture config (encoder name, label vocab sizes, max sequence length)
• tokenizer.json, vocab.txt, tokenizer_config.json, special_tokens_map.json -- DistilBERT tokenizer files

Usage

This checkpoint is loaded via the project's own model.py (not a standard AutoModelForSequenceClassification, since it's a custom joint architecture with two heads). See src/model.py::load_checkpoint and src/predict.py in the VoxParse repo for the loading code.