Epoch	Training Loss	Validation Loss	Accuracy	F1
1	0.928500	0.239526	0.9105	0.9104
2	0.188200	0.160186	0.9300	0.9300
3	0.114900	0.142201	0.9335	0.9349
4	0.089700	0.147365	0.9400	0.9405
5	0.067800	0.159507	0.9330	0.9342
6	0.045800	0.216059	0.9345	0.9350
7	0.029500	0.232973	0.9340	0.9339

Best model was selected at Epoch 4 based on the highest validation F1 score (0.9405). Early stopping triggered after patience of 3 epochs.

Training Hyperparameters

Parameter	Value
Base Model	`distilbert-base-uncased`
Batch Size	64
Learning Rate	5e-5
Weight Decay	0.02
LR Scheduler	Cosine
Warmup Ratio	0.1
Max Epochs	10
Early Stopping Patience	3
FP16	Enabled (on GPU)
Metric for Best Model	F1 (weighted)
Optimizer	AdamW

Dataset

The model was trained on the dair-ai/emotion dataset with a stratified 80/10/10 split to ensure balanced class representation across train, validation, and test sets.

Split	Samples	Percentage
Train	~16,000	80%
Validation	~2,000	10%
Test	~2,000	10%

Usage

Using Pipeline (Recommended)

python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="OmarMaqousi/distilbert-emotion-model-v2",
    return_all_scores=True,
)

result = classifier("I am so happy today!")
print(result)

Using Model Directly

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "OmarMaqousi/distilbert-emotion-model-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "I am so happy today!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)

labels = ["sadness", "joy", "love", "anger", "fear", "surprise"]
for label, score in zip(labels, predictions[0]):
    print(f"{label}: {score:.4f}")

Framework Versions

Transformers: 4.44.2
PyTorch: 2.x
Datasets: latest
Python: 3.12

Limitations

The model is trained on English text only.
Performance may vary on text that is very different from the training data (e.g., formal writing, slang, or domain-specific language).
The model may struggle to distinguish between semantically similar emotions (e.g., sadness vs. fear).