---
library_name: transformers
license: apache-2.0
tags:
  - vietnamese
  - hate-speech-detection
  - phobert
  - bilstm
  - text-classification
pipeline_tag: text-classification
---

# joshswift/phobihsd-proposed

PhoBiHSD proposed checkpoint (PhoBERT-BiLSTM) for Vietnamese hate speech detection on ViHSD.

## Files
- `phobihsd_proposed.pt`: PyTorch checkpoint (`state_dict` + optional thresholds).
- `phobihsd_proposed.meta.json`: model metadata.
- `model_comparison.yaml`: training/inference config used by this repository.

## Label Mapping
- `0`: Clean
- `1`: Offensive
- `2`: Hate

## Metadata (from checkpoint sidecar)
```json
{
  "model_name": "vinai/phobert-base-v2",
  "head_type": "cls_mlp",
  "hidden_dim": 256,
  "dropout": 0.5,
  "freeze_encoder": false,
  "max_len": 100,
  "thresholds": [
    0.4,
    0.4,
    0.6
  ],
  "seed": 4
}
```

## Inference (from this repo)
```bash
export PHOBIHSD_PROPOSED_CKPT=phobihsd_proposed.pt
export PHOBIHSD_CONFIG=config/experiments/model_comparison.yaml
python -m app.gradio_app
```