logasanjeev/goemotions-bert

GoEmotions BERT Classifier

Fine-tuned BERT-base-uncased on go_emotions for multi-label classification (28 emotions).

Model Details

• Architecture: BERT-base-uncased (110M parameters)
• Training Data: GoEmotions (58k Reddit comments, 28 emotions)
• Loss Function: Focal Loss (gamma=2)
• Optimizer: AdamW (lr=2e-5, weight_decay=0.01)
• Epochs: 5
• Hardware: Kaggle T4 x2 GPUs

Try It Out

For accurate predictions with optimized thresholds, use the Gradio demo.

Performance

• Micro F1: 0.6025 (optimized thresholds)
• Macro F1: 0.5266
• Precision: 0.5425
• Recall: 0.6775
• Hamming Loss: 0.0372
• Avg Positive Predictions: 1.4564

Class-Wise Performance

The following table shows per-class metrics on the test set using optimized thresholds (see thresholds.json):

Emotion	F1 Score	Precision	Recall	Support
admiration	0.7022	0.6980	0.7063	504
amusement	0.8171	0.7692	0.8712	264
anger	0.5123	0.5000	0.5253	198
annoyance	0.3820	0.2908	0.5563	320
approval	0.4112	0.3485	0.5014	351
caring	0.4601	0.4045	0.5333	135
confusion	0.4488	0.4533	0.4444	153
curiosity	0.5721	0.4402	0.8169	284
desire	0.4068	0.6857	0.2892	83
disappointment	0.3476	0.3220	0.3775	151
disapproval	0.4126	0.3433	0.5169	267
disgust	0.4950	0.6329	0.4065	123
embarrassment	0.5000	0.7368	0.3784	37
excitement	0.4084	0.4432	0.3786	103
fear	0.6311	0.5078	0.8333	78
gratitude	0.9173	0.9744	0.8665	352
grief	0.2500	0.5000	0.1667	6
joy	0.6246	0.5798	0.6770	161
love	0.8110	0.7630	0.8655	238
nervousness	0.3830	0.3750	0.3913	23
optimism	0.5777	0.5856	0.5699	186
pride	0.4138	0.4615	0.3750	16
realization	0.2421	0.5111	0.1586	145
relief	0.5385	0.4667	0.6364	11
remorse	0.6797	0.5361	0.9286	56
sadness	0.5391	0.6900	0.4423	156
surprise	0.5724	0.5570	0.5887	141
neutral	0.6895	0.5826	0.8444	1787

Usage

The model uses optimized thresholds stored in thresholds.json for predictions. Example in Python:

from transformers import BertForSequenceClassification, BertTokenizer
import torch
import json
import requests

# Load model and tokenizer
repo_id = "logasanjeev/goemotions-bert"
model = BertForSequenceClassification.from_pretrained(repo_id)
tokenizer = BertTokenizer.from_pretrained(repo_id)

# Load thresholds
thresholds_url = f"https://huggingface.co/{repo_id}/raw/main/thresholds.json"
thresholds_data = json.loads(requests.get(thresholds_url).text)
emotion_labels = thresholds_data["emotion_labels"]
thresholds = thresholds_data["thresholds"]

# Predict
text = "I’m just chilling today."
encodings = tokenizer(text, padding='max_length', truncation=True, max_length=128, return_tensors='pt')
with torch.no_grad():
    logits = torch.sigmoid(model(**encodings).logits).numpy()[0]
predictions = [(emotion_labels[i], logit) for i, (logit, thresh) in enumerate(zip(logits, thresholds)) if logit >= thresh]
print(sorted(predictions, key=lambda x: x[1], reverse=True))
# Output: [('neutral', 0.8147)]