Fine-tuned ai-forever/sbert_large_nlu_ru for multi-label emotion detection in Russian text. Trained on CEDR v1 with Focal Loss and per-class decision thresholds.

Metrics

Metric	Value
F1 macro	0.7826
F1 micro	0.8183
F1 weighted	0.8174
Precision micro	0.8025
Recall micro	0.8348

Per-class F1 (with per-class thresholds)

Class	F1	Threshold
joy	0.8753	0.56
sadness	0.8642	0.59
fear	0.7841	0.69
surprise	0.7563	0.71
anger	0.6329	0.69

Usage

python

import json, torch, requests
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "ilyali034/russian-emotion-classifier-sbert-large"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

cfg = json.loads(
    requests.get(
        f"https://huggingface.co/{model_id}/resolve/main/emotion_config.json"
    ).text
)
labels_list = cfg["labels"]
thresholds  = cfg["thresholds"]

def predict(text: str) -> dict:
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        probs = torch.sigmoid(model(**inputs).logits)[0].cpu().numpy()
    return {
        lbl: round(float(p), 3)
        for lbl, p in zip(labels_list, probs)
        if p > thresholds[lbl]
    }

predict("Я очень рад, но немного боюсь")
# {'joy': 0.912, 'fear': 0.701}

Training details

Parameter	Value
Base model	ai-forever/sbert_large_nlu_ru (426M params)
Dataset	sagteam/cedr_v1 (7528 train / 1882 test)
Loss	Focal BCE (γ=1.0) + pos_weight per class
Label smoothing	0.1
Effective batch size	32 (batch=8 × grad_accum=4)
Learning rate	5e-6 (cosine schedule, warmup 10%)
Epochs	3
Thresholds	Per-class, optimized on validation set
Hardware	Tesla T4 (~17 min)

Why Focal Loss: The dataset is heavily imbalanced — anger appears 5.5× less often than joy (411 vs 1569 examples). Focal Loss with γ=1.0 down-weights easy examples and keeps training focused on rare classes.

Why per-class thresholds: All classes received optimal thresholds above 0.5 (range 0.56–0.71), reflecting the model's tendency to underestimate confidence under label smoothing. Thresholds are stored in emotion_config.json and must be loaded for correct inference (see usage example above).

Labels

joy · sadness · surprise · fear · anger

Citation

If you use this model, please cite the dataset:

bibtex

@dataset{cedr_v1,
  author = {SAGTeam},
  title  = {CEDR: Russian Emotion Dataset},
  year   = {2023},
  url    = {https://huggingface.co/datasets/sagteam/cedr_v1}
}