This model is a fine-tuned version of answerdotai/ModernBERT-base for multilabel classification of Motivational Interviewing (MI) behavioral codes. It classifies utterances into three non-mutually-exclusive categories used in behavioral coding of therapeutic conversations.

Developed by: Lekhansh

Model type: Multilabel Text Classification

Language: English

Base model: answerdotai/ModernBERT-base

License: Apache 2.0

Intended Uses

Primary Use Case

This model is designed for automated behavioral coding in Motivational Interviewing contexts, predicting three types of MI-consistent and MI-inconsistent behaviors:

Adherent: MI-adherent behaviors (e.g., affirmations, seek collaboration)
Non-Adherent: MI-non-adherent behaviors (e.g., confrontation, persuade without permission)
Neutral: Neutral behaviors (e.g., giving information, questions, reflections)

Key Features

Multilabel Classification: Utterances can have multiple labels simultaneously
Therapeutic Context: Specifically trained on Motivational Interviewing conversations
Context-Aware: Includes three preceding utterances for context

Potential Applications

Automated analysis of therapy session transcripts
Training and feedback for MI practitioners
Quality assurance in behavioral health interventions
Research in therapeutic communication patterns

Model Performance

Test Set Metrics

The model was evaluated on a held-out test set of 3,235 coded utterances.

Overall Performance

Metric	Score
Exact Match Accuracy	85.63%
Hamming Loss	0.0579
F1 Macro	86.66%
F1 Micro	92.46%
Precision Macro	86.53%
Precision Micro	93.47%
Recall Macro	86.84%
Recall Micro	91.48%

Exact Match: Percentage of examples where all labels are predicted correctly Hamming Loss: Average fraction of labels that are incorrectly predicted (lower is better)

Per-Label Performance

Label	F1 Score	Precision	Recall	Accuracy
Adherent	74.29%	74.47%	74.10%	90.26%
Non-Adherent	89.32%	87.34%	91.39%	98.98%
Neutral	96.39%	97.77%	95.04%	93.38%

Class Distribution

The training data exhibits class imbalance, addressed through positive class weighting:

Neutral: Most common (majority class)
Non-Adherent: Moderate frequency
Adherent: Least common (minority class)

Training Details

Training Data

Source: Multilabel behavioral coding dataset from Motivational Interviewing transcripts
Preprocessing:
- Excluded utterances marked as "not_coded" (no MI codes assigned)
- Included context from three preceding utterances
- Stratified splitting to maintain label distribution
Split: 70% train, 15% validation, 15% test

Training Procedure

Hardware:

GPU training with CUDA
Mixed precision (BFloat16) training

Hyperparameters:

Parameter	Value
Learning Rate	6e-5
Batch Size (per device)	12
Gradient Accumulation	2 steps
Effective Batch Size	24
Max Sequence Length	3000 tokens
Epochs	20 (early stopped at epoch 14)
Weight Decay	0.01
Warmup Ratio	0.1
LR Scheduler	Cosine
Optimizer	AdamW
Dropout	0.1

Training Features:

Positive Class Weighting: BCEWithLogitsLoss with computed pos_weights for each label
Early Stopping: Patience of 3 epochs on validation F1 macro
Gradient Checkpointing: Enabled for memory efficiency
Flash Attention 2: For efficient attention computation
Best Model Selection: Based on validation F1 macro score

Loss Function: Binary Cross-Entropy with Logits Loss (BCEWithLogitsLoss) with per-label positive class weights

Model Architecture

The model uses a custom architecture on top of ModernBERT:

ModernBERT-base (encoder)
  → [CLS] token extraction
  → Dropout (0.1)
  → Linear layer (hidden_size → 3)
  → Sigmoid activation (applied during inference)

Usage

Direct Use

python

import torch
from transformers import AutoTokenizer, AutoModel
import torch.nn as nn

# Define the model class
class MultiLabelBERTModel(nn.Module):
    def __init__(self, model_name, num_labels=3, dropout=0.1):
        super().__init__()
        self.bert = AutoModel.from_pretrained(model_name)
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(self.bert.config.hidden_size, num_labels)
        self.num_labels = num_labels

    def forward(self, input_ids, attention_mask):
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = outputs.last_hidden_state[:, 0, :]  # [CLS] token
        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        return logits

# Load model and tokenizer
model_name = "Lekhansh/bc-multilabel-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Initialize model architecture
model = MultiLabelBERTModel(model_name, num_labels=3)

# Load trained weights
# Note: You'll need to load the weights from the saved model
model.eval()

# Prepare input
text = "That's a wonderful goal you've set for yourself."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=3000)

# Get predictions
with torch.no_grad():
    logits = model(inputs['input_ids'], inputs['attention_mask'])
    probs = torch.sigmoid(logits)
    predictions = (probs > 0.5).int()

# Interpret results
labels = ['adherent', 'non_adherent', 'neutral']
print(f"Text: {text}")
print("\nPredictions:")
for i, label in enumerate(labels):
    if predictions[0][i]:
        print(f"  ✓ {label} (confidence: {probs[0][i]:.2%})")

Batch Prediction with Confidence Scores

python

def predict_multilabel(texts, model, tokenizer, threshold=0.5):
    """
    Predict multiple labels for each text with confidence scores.

    Args:
        texts: List of input texts
        model: The multilabel classification model
        tokenizer: The tokenizer
        threshold: Probability threshold for positive prediction (default: 0.5)

    Returns:
        List of dicts with predictions and probabilities
    """
    inputs = tokenizer(
        texts,
        return_tensors="pt",
        truncation=True,
        max_length=3000,
        padding=True
    )

    with torch.no_grad():
        logits = model(inputs['input_ids'], inputs['attention_mask'])
        probs = torch.sigmoid(logits)

    labels = ['adherent', 'non_adherent', 'neutral']
    results = []

    for i in range(len(texts)):
        predictions = (probs[i] > threshold).int()
        result = {
            'text': texts[i],
            'labels': {},
            'probabilities': {}
        }

        for j, label in enumerate(labels):
            result['labels'][label] = bool(predictions[j])
            result['probabilities'][label] = float(probs[i][j])

        results.append(result)

    return results

# Example usage
utterances = [
    "I hear you saying that you want to change but you're not sure how.",
    "You need to stop making excuses and just do it.",
    "How many cigarettes do you smoke per day?"
]

results = predict_multilabel(utterances, model, tokenizer)
for r in results:
    print(f"\nText: {r['text'][:60]}...")
    print("Predicted labels:")
    for label in ['adherent', 'non_adherent', 'neutral']:
        status = "✓" if r['labels'][label] else "✗"
        conf = r['probabilities'][label]
        print(f"  {status} {label}: {conf:.2%}")

Custom Threshold Tuning

python

# Adjust threshold for precision/recall trade-off
def predict_with_custom_threshold(text, model, tokenizer, thresholds):
    """
    Predict with different thresholds for each label.

    Args:
        thresholds: Dict with keys 'adherent', 'non_adherent', 'neutral'
    """
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=3000)

    with torch.no_grad():
        logits = model(inputs['input_ids'], inputs['attention_mask'])
        probs = torch.sigmoid(logits)

    labels_list = ['adherent', 'non_adherent', 'neutral']
    predictions = {}

    for i, label in enumerate(labels_list):
        threshold = thresholds.get(label, 0.5)
        predictions[label] = {
            'predicted': bool(probs[0][i] > threshold),
            'probability': float(probs[0][i]),
            'threshold': threshold
        }

    return predictions

# Example: Higher threshold for adherent (higher precision)
custom_thresholds = {
    'adherent': 0.6,
    'non_adherent': 0.5,
    'neutral': 0.5
}

result = predict_with_custom_threshold(
    "What are your thoughts on reducing your drinking?",
    model,
    tokenizer,
    custom_thresholds
)

Limitations and Bias

Limitations

Domain Specificity: Trained on Motivational Interviewing data; may not generalize to other therapeutic modalities
Context Dependency: Performance may vary with utterances lacking proper conversational context
Class Imbalance: Lower performance on "adherent" label due to class imbalance in training data
Multilabel Complexity: Some utterances may have ambiguous or overlapping codes
Context Length: Maximum 3000 tokens; longer texts will be truncated
Language: Trained on English text only

Potential Biases

Training data may reflect biases from the original coding framework and human coders
Performance may vary across different MI contexts (e.g., substance use vs. health behavior change)
Cultural and linguistic variations in therapeutic communication may affect predictions
The model may be more accurate on populations/contexts similar to training data

Recommended Use

Use as a screening tool or preliminary analysis, not as definitive behavioral coding
Validate predictions with human expert review, especially for critical applications
Consider adjusting prediction thresholds based on your use case (precision vs. recall trade-off)
Be aware that multilabel predictions may sometimes conflict with clinical judgment

Technical Specifications

Model Architecture

Base: ModernBERT-base (encoder-only transformer)
Custom Head: Dropout (0.1) + Linear layer (hidden_size → 3 labels)
Activation: Sigmoid (for independent label probabilities)
Attention: Flash Attention 2 implementation
Parameters: ~110M (inherited from base model + classification head)
Precision: BFloat16

Compute Infrastructure

Training: Single GPU with CUDA
Inference: CPU or GPU compatible
Memory: ~500MB model size

Label Format

python

# Output format
{
  "adherent": 0 or 1,
  "non_adherent": 0 or 1,
  "neutral": 0 or 1
}

# Example: An utterance can have multiple labels
# "I hear that you're struggling, and I believe you can overcome this."
# → adherent=1, non_adherent=0, neutral=0

Environmental Impact

Training was conducted using mixed precision to optimize resource usage. Exact carbon footprint was not measured.

Citation

If you use this model in your research, please cite:

bibtex

@misc{lekhansh2025bcmultilabel,
  author = {Lekhansh},
  title = {Behavioral Coding Multilabel Classifier for Motivational Interviewing},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/Lekhansh/bc-multilabel-classifier}}
}

References

For more information on Motivational Interviewing behavioral coding:

Miller, W. R., & Rollnick, S. (2013). Motivational Interviewing: Helping People Change (3rd ed.)
Moyers, T. B., et al. (2016). Motivational Interviewing Treatment Integrity Coding Manual 4.2.1

Model Card Authors

Lekhansh

Model Card Contact

[drlekhansh@gmail.com]