vad-bert / README.md
Psycheros's picture
Upload folder using huggingface_hub
e6f5509 verified
|
Raw History Blame Contribute Delete
1.99 kB
---
language: en
tags:
- vad
- emotion
- bert
- onnx
license: mit
base_model:
- RobroKools/vad-bert
datasets:
- reallycarlaost/emobank
---
# vad-bert (ONNX + int8, Psycheros builds)
ONNX exports of [RobroKools/vad-bert](https://huggingface.co/RobroKools/vad-bert)
(BERT-base fine-tuned on EmoBank for Valence/Arousal/Dominance regression),
prepared for [transformers.js](https://huggingface.co/docs/transformers.js)-style
in-process inference — no PyTorch, no server.
Two variants, identical behavior:
| variant | size | notes |
|---|---|---|
| `int8/` | ~110MB | dynamic int8 quantization; ~2–2.6× faster on CPU, ~0.02 delta drift |
| `fp32/` | ~438MB | bit-exact conversion of the original weights |
Each folder is a standalone transformers.js layout: tokenizer/config at the
root, `onnx/model.onnx` single-file. In both variants the model file is named
`onnx/model.onnx` regardless of precision.
## What was done to the weights
- `fp32/` — ONNX export of the original `pytorch_model.bin` via optimum
(format conversion only; weights unchanged, verified output-identical).
- `int8/` — dynamic int8 quantization of that export via onnxruntime
(`quantize_dynamic`, QInt8).
- Tokenizer and config files are re-serialized by the export tooling —
semantically identical (same tokens, same ids, same model config), but the
JSON shape differs from the original repo's (newer transformers serializes
special tokens in expanded form). `tokenizer.json` and `vocab.txt` are
byte-identical to the original.
## Credits & license
- Original model: [RobroKools/vad-bert](https://huggingface.co/RobroKools/vad-bert) — MIT
- Base model: [google-bert/bert-base-uncased](https://huggingface.co/google-bert/bert-base-uncased) — Apache 2.0
- Training data: [EmoBank](https://github.com/JULIELab/EmoBank) (via reallycarlaost/emobank)
This repository is distributed under the same MIT license as the original;
see [LICENSE](./LICENSE) and the original's license above.