---
language: ru
license: apache-2.0
library_name: transformers
pipeline_tag: text-classification
base_model: Aniemore/rubert-tiny2-russian-emotion-detection
base_model_relation: quantized
datasets:
- Aniemore/cedr-m7
tags:
- text-classification
- emotion-recognition
- russian
- multi-label-classification
- quantized
- compressed-tensors
- int8
- fp8
- int4
metrics:
- roc_auc
- f1
- accuracy
model-index:
- name: rubert-tiny2-russian-emotion-detection-quantized
  results:
  - task:
      name: Emotion Recognition
      type: text-classification
    dataset:
      name: CEDR-m7 test (int8)
      type: Aniemore/cedr-m7
      args: ru
    metrics:
    - name: ROC AUC macro (int8)
      type: roc_auc
      value: 0.8696
    - name: Macro F1 (int8)
      type: f1
      value: 0.6013
  - task:
      name: Emotion Recognition
      type: text-classification
    dataset:
      name: CEDR-m7 test (fp8)
      type: Aniemore/cedr-m7
      args: ru
    metrics:
    - name: ROC AUC macro (fp8)
      type: roc_auc
      value: 0.8695
    - name: Macro F1 (fp8)
      type: f1
      value: 0.6014
  - task:
      name: Emotion Recognition
      type: text-classification
    dataset:
      name: CEDR-m7 test (int4)
      type: Aniemore/cedr-m7
      args: ru
    metrics:
    - name: ROC AUC macro (int4)
      type: roc_auc
      value: 0.8738
    - name: Macro F1 (int4)
      type: f1
      value: 0.5981
---

# rubert-tiny2-russian-emotion-detection · quantized

Quantized builds of [`Aniemore/rubert-tiny2-russian-emotion-detection`](https://huggingface.co/Aniemore/rubert-tiny2-russian-emotion-detection) — multi-label emotion recognition for Russian text over seven classes: `anger`, `disgust`, `enthusiasm`, `fear`, `happiness`, `neutral`, `sadness`.

The weights here are the published original, quantized. They were not retrained and they are not a different model.

## Variants

| subfolder | scheme | weights | ROC AUC (macro) | macro-F1 | WA | UA |
|---|---|---:|---:|---:|---:|---:|
| _(original repo)_ | fp32 | 111 MiB | 0.8696 | 0.6015 | 0.7662 | 0.6050 |
| `int8` | W8A16 | 107 MiB | 0.8696 | 0.6013 | 0.7662 | 0.6050 |
| `fp8` | W8A16-float | 107 MiB | 0.8695 | 0.6014 | 0.7646 | 0.6028 |
| `int4` | W4A16_ASYM | 106 MiB | 0.8738 | 0.5981 | 0.7588 | 0.6041 |

<img src="assets/quality.svg" alt="Quality after quantization" width="760">

<img src="assets/size.svg" alt="Weights on disk" width="760">

ROC AUC is listed first because the head is multi-label: macro-F1 depends on the decision threshold, which is 0.5 here because that is what the head was trained under, while ROC AUC does not.

### How much this actually saves

Only `Linear` layers are quantized. In a BERT classifier the embedding matrix is not one of them, and on the smaller models it is most of the checkpoint &mdash; so the saving here scales with the encoder rather than with the parameter count. The large model compresses well; `rubert-tiny` barely moves, and the table above says so rather than quoting a ratio from the layers that did shrink.

## Usage

```python
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "Aniemore/rubert-tiny2-russian-emotion-detection-quantized"
model = AutoModelForSequenceClassification.from_pretrained(
    repo, subfolder="int8").eval()          # or "fp8", "int4"
tok = AutoTokenizer.from_pretrained(repo, subfolder="int8")

x = tok("мне сегодня очень грустно", return_tensors="pt")
with torch.no_grad():
    # multi-label: sigmoid per class, not softmax over classes
    probs = model(**x).logits.sigmoid()[0]
print({model.config.id2label[i]: round(p.item(), 3) for i, p in enumerate(probs)})
```

## Limitations

- Weight-only, round-to-nearest, no calibration.
- Scored on the CEDR-m7 test split only. CEDR is written text; performance on transcribed speech, which carries no punctuation and no casing, is not measured here.
- Inherited from [`cointegrated/rubert-tiny2`](https://huggingface.co/cointegrated/rubert-tiny2); the licence follows the base model.