This model is a fine-tuned version of microsoft/mdeberta-v3-base for multi-label emotion recognition in social media data. It was trained to handle 6 Ekman's emotions (anger, disgust, fear, joy, sadness, surprise) and a neutral category. It supports both Russian and English texts.

Model training

Public emotion recognition datasets were used:

GoEmotions, English, all 7 classes;
CEDR, Russian, lacks disgust;
BRIGHTER, Russian and English; Russian has all 7 classes, while English lacks disgust.

Training was done with the following:

LoRA fine-tuning (r=16; a=32; q, k, v, dense projections; dropout=0.1);
Clipped asymmetric loss with epsilon=0.04, gamma_neg=3.5 and gamma_pos=1.5 to handle the class imbalance;
AdamW as an optimizer with LR=2e-4 for LoRA and LR=1e-5, epsilon=1e-3 for classification head;
Linear learning rate scheduler with 10% warmup steps;
Batch size of 16 with 2-step gradient accumulation;
Torch AMP and gradient norm clipping of 1.0;
10 scheduled epochs, with training stopping early after epoch 7 due to increasing validation loss.

Evaluation

Model achieved 0.9291 macro AUC and 0.6957 macro F1 for 7-class multilabel emotion recognition on the combined GoEmotion + CEDR + BRIGHTER set. F1 was calculated with validation-optimal thresholds. Per-emotion breakdown is shown below:

Emotion	AUC	F1
Anger	0.9113	0.6198
Fear	0.9505	0.7717
Joy	0.9376	0.7937
Sadness	0.9245	0.7043
Disgust	0.9725	0.6442
Surprise	0.9048	0.6432
Neutral	0.9024	0.6931

How to Use

You can use pipeline with top_k=None and function_to_apply='sigmoid' to get probabilities for all classes with a simple callable API.

python

from transformers import pipeline

model_path = "catgamer1/mdeberta-ru-en-emotion"

classifier = pipeline(
    "text-classification",
    model=model_path,
    tokenizer=model_path,
    top_k=None, 
    function_to_apply='sigmoid'
)

sentence = "What a relief, the tire is fine!"
result = classifier(sentence)