🔍 Token Log Probability Analyzer - Model Comparison

Compare how two ERNIE models predict each token in your text with detailed log probability breakdown.

ERNIE-4.5-Base-PT

Token Analysis

ERNIE-4.5-PT

Token Analysis

Try these examples:

How to Interpret Results

This interface compares two ERNIE models side by side:

  1. ERNIE-4.5-Base-PT (left): Base model, better at general language patterns
  2. ERNIE-4.5-PT (right): Instruction-tuned model, better at following complex instructions

Analysis Components

For each model, you'll see:

  • Summary: Key metrics including Total Log Probability and average token probability
  • Token Analysis Table: Detailed breakdown of each token's log probability and probability
  • Token Probability Chart: Visual representation of each token's prediction probability

Model Comparison

  • Model Comparison Summary: Shows which model has higher overall confidence
  • Model Comparison Chart: Side-by-side visualization of token probabilities

Key Concepts

  • Log Probability:

    • Ranges from -∞ to 0
    • Closer to 0 = higher model confidence
    • Used instead of raw probability to avoid numerical underflow
  • Total Log Probability:

    • Sum of individual token log probabilities
    • Measures overall model confidence in the entire sequence
    • Allows comparison between different models
  • Why Compare Models?:

    • Base models may be better at general language
    • Instruction-tuned models may be better at specific tasks
    • Different models have different strengths for different types of text