RoBERTa-base Fine-tuned on GoEmotions (28 classes)
This model is a roberta-base encoder fine-tuned on the GoEmotions dataset for 28-class multi-label emotion classification.
Produced as the final classifier (Ablation 6) for a bachelor's thesis: LLM-Based Emotion-Aware Adaptive Learning Assistant. Serves as the emotion detection component of a CVT-grounded tutoring pipeline.
Training Details
| Parameter | Value |
|---|---|
| Base model | roberta-base |
| Dataset | GoEmotions — standard train/val/test split |
| Loss | Clipped Asymmetric Loss (γ_pos=1, γ_neg=3, clip=0.04) |
| Optimizer | AdamW + Layer-wise LR Decay (decay=0.9, top_lr=2e-5) |
| Scheduler | Cosine annealing with 10% warmup |
| Batch size | 32 |
| Best epoch | 5 of 10 (early stopping patience=3) |
| Threshold | 0.5 (per-label sigmoid) |
Test Set Results
| Metric | Score |
|---|---|
| Macro-F1 | 0.5472 |
| Micro-F1 | 0.6119 |
| Precision (macro) | 0.5293 |
| Recall (macro) | 0.5930 |
| Zero-F1 classes | 0 |
Generalisation ratio (test / val macro-F1): 1.010
Labels (index 0-27)
admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise, neutral
Inference
from transformers import RobertaTokenizerFast, RobertaForSequenceClassification
import torch
tokenizer = RobertaTokenizerFast.from_pretrained("hoorSobh/roberta_ablation6")
model = RobertaForSequenceClassification.from_pretrained("hoorSobh/roberta_ablation6")
model.eval()
text = "I do not understand this at all"
inputs = tokenizer(text, return_tensors='pt', truncation=True, max_length=128)
with torch.no_grad():
probs = torch.sigmoid(model(**inputs).logits).squeeze()
threshold = 0.5
detected = [model.config.id2label[i] for i, p in enumerate(probs.tolist()) if p >= threshold]
top1 = model.config.id2label[probs.argmax().item()]
print("Top-1 emotion:", top1)
print("All detected: ", detected)
Citation
Demszky et al. (2020). GoEmotions: A Dataset of Fine-Grained Emotions. ACL 2020.