---
license: other
library_name: pytorch
tags:
  - emotion-recognition
  - distillation
  - efficientnet
  - multitask
datasets:
  - aussiegingersnap2/scroll-happy-emotion
metrics:
  - spearmanr
model-index:
  - name: v0.3e-bias-w1p5-seed-101
    results:
      - task:
          type: image-classification
          name: Soft-target Emotion Distillation
        dataset:
          name: scroll-happy-emotion (training_faces, val split)
          type: aussiegingersnap2/scroll-happy-emotion
        metrics:
          - type: spearmanr
            name: Mean row Spearman (emotions)
            value: 0.7981
          - type: spearmanr
            name: Mean row Spearman (FACS)
            value: 0.8919
          - type: spearmanr
            name: Mean row Spearman (descriptions)
            value: 0.6896
          - type: accuracy
            name: Top-1 emotion vs teacher argmax
            value: 0.3678
---

# Scroll Happy Emotion — Student (efficientnet_b2)

EfficientNet student distilled from Hume teacher labels on
[`aussiegingersnap2/scroll-happy-emotion`](https://huggingface.co/datasets/aussiegingersnap2/scroll-happy-emotion).

Predicts continuous probabilities for **48 emotions**, **36 FACS AUs**, and
**27 facial descriptions** from a single face crop.

## Eval (held-out creators)

| Metric | Value |
|---|---|
| Mean row Spearman (emotions) | 0.7981 |
| Mean row Spearman (FACS) | 0.8919 |
| Mean row Spearman (descriptions) | 0.6896 |
| Top-1 emotion vs teacher | 0.3678 |
| Top-3 emotion recall vs teacher | 0.6356 |
| Dead emotion dims (val pred std < 1e-3) | 0/48 |
| Val rows | 15,896 |

Trackio run: `v0.3e-bias-w1p5-seed-101` in project `scroll-happy-emotion`.