import gradio as gr
from typing import Tuple
def render_score_card(
ttr_output: float,
ttr_train: float,
similarity: float,
mad_score: float,
risk_level: str,
) -> Tuple[str, str, str]:
"""Renders the evaluation summary and score details."""
color = {"High": "#e57373", "Medium": "#ffb74d", "Low": "#81c784"}[risk_level]
risk_explanations = {
"High": """
π¨ **High Risk Detected** Your model outputs are **very similar** to your planned training data.
This suggests a **strong feedback loop**, meaning the model is likely to reinforce existing patterns rather than learning new behaviors.
**What You Can Do**:
- Replace synthetic data with more **diverse real user input** - Use **paraphrasing techniques** before reuse
- Add **augmentation or filtering** before retraining
""",
"Medium": """
π **Moderate Risk Identified** There is some overlap between your outputs and training content.
Your model may partially reinforce existing phrasing patterns.
**Suggestions**:
- Mix synthetic and real inputs carefully
- Monitor training logs for semantic redundancy
""",
"Low": """
π’ **Low Risk Score** Your model output and training data appear **diverse** and distinct.
This is a good sign that your model is learning from **new and varied sources**.
**Youβre on the right track!**
""",
}
summary = f"""
### π Evaluation Summary
**Lexical Diversity (Output):** {ttr_output:.2f}
TTR = unique words / total words
**Lexical Diversity (Training Set):** {ttr_train:.2f}
Broader content = higher TTR
**Semantic Similarity (Cosine):** {similarity:.2f}
Cosine similarity between embeddings
MAD Risk Score: {mad_score:.2f} β {risk_level} Risk
"""
details = f"""
π Score Breakdown
TTR Component (0.3 Γ (1 - TTR)): {(1 - ttr_output) * 0.3:.2f}
Similarity Component (0.7 Γ Cosine): {similarity * 0.7:.2f}
MAD Score = 0.3 Γ (1 - TTR) + 0.7 Γ Semantic Similarity
"""
explanation = f"""
π What does this score mean?
{risk_explanations[risk_level]}
"""
return summary, details, explanation