my_lead_model

DistilBERT Sales Lead Classification Model (v2 — Domain Adapted)

This model is a fine-tuned version of distilbert-base-uncased, trained in two stages to score sales leads by sentiment for CRM lead prioritization.

Training Pipeline

Stage 1 — Base Fine-Tune: Trained on the Financial PhraseBank dataset (sentences_75agree config) for 3-class sentiment classification (negative / neutral / positive), replacing an earlier Rotten Tomatoes movie-review proxy with business-domain text.

Stage 2 — Sales Domain Adaptation: Continued training on a 240-example, hand-authored dataset of sales/CRM lead interaction sentences (80 per class), covering patterns the base model struggled with — negation ("not interested"), competitor retention ("happy with current vendor"), budget objections, and clear buying signals. Trained with a low learning rate (2e-5) to adapt without catastrophic forgetting of the base.

Training Results

Stage 1 (Financial PhraseBank, 2 epochs):

Stage 2 (Sales domain adaptation, 3 epochs):

Final model achieves 85.4% accuracy on held-out sales-domain sentences.

Labels

• 0 — Negative (lost lead / not interested / competitor retention)
• 1 — Neutral (hedging / evaluating / no decision yet)
• 2 — Positive (buying signal / ready to move forward)

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="kaizen696/my_lead_model"
)

# Example: High Quality
print(classifier("Budget is approved, excited to move forward."))

# Example: Low Quality
print(classifier("Not interested at all, too expensive."))

Known Limitations

• Trained on a small (240-example) domain-adaptation set — may still misclassify sentences with mixed lexical signals (e.g. positive-sounding words like "happy" paired with a negative outcome, such as "happy with their current vendor").
• Financial PhraseBank base introduces some finance-specific vocabulary bias not fully corrected by the sales fine-tune.
• Recommended as a lead-prioritization signal, not a sole decision-making input.