enhanced-bert-coqa-multilabel-classifier

Enhanced BERT Multi-Label Classifier for Conversational QA

Performance

• Micro F1: 0.7779
• Macro F1: 0.7098
• Optimal Threshold: 0.20 (CRITICAL - not 0.5!)

Class Performance

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained("IIchukissII/enhanced-bert-coqa-multilabel-classifier")
tokenizer = AutoTokenizer.from_pretrained("IIchukissII/enhanced-bert-coqa-multilabel-classifier")

text = "context [SEP] question [SEP] answer"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits)

# IMPORTANT: Use threshold 0.20, not 0.5!
LABELS = ["interactive", "responsive", "questioning", "collaborative"]
predictions = {label: int(probs[0, i] >= 0.20) for i, label in enumerate(LABELS)}

Architecture

• Enhanced BERT with 3-layer classifier head
• Layer normalization and L2 regularization
• Optimized for multi-label classification

Training

• 7 epochs on expanded dataset
• Breakthrough in interactive detection (+15.7% F1)
• Threshold optimization discovery (0.20 optimal)