Animhaven/deberta-v1-base-russian-go-emotions-ONNX

deberta-v1-base-russian-go-emotions (INT8 ONNX)

INT8-quantized ONNX export of fyaronskiy/deberta-v1-base-russian-go-emotions for browser-side inference via Transformers.js.

• 28 GoEmotions classes (multi-label)
• ~125M parameters (deBERTa-v1-base)
• ~120 MB INT8 ONNX (vs 476 MB FP32 — 4× smaller)
• ~50 ms latency on a typical laptop CPU

Use with Transformers.js

import { pipeline } from "@huggingface/transformers";

const classify = await pipeline(
  "text-classification",
  "KonstantinSerg/deberta-v1-base-russian-go-emotions-ONNX",
  { quantized: true, top_k: null }, // top_k=null → return all 28 scores
);

const result = await classify("Я в полном восторге, наконец-то у меня получилось!");
// → [{ label: "excitement", score: 0.78 }, { label: "joy", score: 0.52 }, ...]

Quality (35 hand-curated Russian sentences)

Metric	Score
Recall@1 (exact 28-class label)	97% (34/35)
Recall@3	97%
10-class avatar match (mapped via GoEmotions → avatar emotions)	100%

INT8 vs FP32 PyTorch agreement: 5/5 on a representative subset, no observed quality regression from quantization.

Labels (id2label)

0 admiration   1 amusement     2 anger        3 annoyance
4 approval     5 caring        6 confusion    7 curiosity
8 desire       9 disappointment 10 disapproval 11 disgust
12 embarrassment 13 excitement 14 fear         15 gratitude
16 grief       17 joy           18 love        19 nervousness
20 optimism    21 pride         22 realization 23 relief
24 remorse     25 sadness       26 surprise    27 neutral

Files

• onnx/model_quantized.onnx (120 MB) — INT8-quantized weights, the one to load
• tokenizer.json (5 MB) — ByteLevel BPE tokenizer (RoBERTa/deBERTa-v1 style)
• config.json — model config with id2label
• merges.txt, vocab.json — tokenizer support files

License

Inherits Apache-2.0 from the upstream model.

Citation

Original model:

fyaronskiy/deberta-v1-base-russian-go-emotions
https://huggingface.co/fyaronskiy/deberta-v1-base-russian-go-emotions