agusdac/yt-comment-purchase-intent-roberta

Purchase Intent Lead Scorer (DistilRoBERTa-base)

This model is a fine-tuned version of distilroberta-base trained on a specialized dataset of over 600+ user comments. It is designed to distinguish between casual chatter and potential commercial interest.

Model Details

• Base Model: distilroberta-base
• Task: Binary Text Classification (No Intent vs. Intent)
• Training Set: 600+ manually labeled web comments and social media interactions.
• Efficiency: DistilRoBERTa is ~2x faster and 33% smaller than standard RoBERTa, making it ideal for real-time lead scoring in TypeScript/Node.js environments.

Classification Logic

The model outputs two primary labels:

1. no intent: General inquiry, noise, or non-commercial chatter.
2. intent: Expressed desire to purchase, request for pricing, or product availability questions.

Lead Scoring Thresholds

For production use, we recommend the following logic based on the intent score:

• Hot Lead (Explicit): Score ≥ 0.98
• Warm Lead (Implicit): Score 0.70 - 0.97
• Potential Interest: Score < 0.70

How to use (Inference API)

const MODEL_ID = "agusdac/yt-comment-purchase-intent-roberta";

async function checkLead(text: string) {
  const response = await fetch(`https://router.huggingface.co/hf-inference/models/${MODEL_ID}`, {
    method: "POST",
    headers: { Authorization: `Bearer ${process.env.HF_TOKEN}` },
    body: JSON.stringify({ inputs: text }),
  });
  return await response.json();
}