Expand description
Quality Filter Module
Provides heuristic-based output filtering using generation metrics.
§Important Distinction
This module is for filtering outputs, NOT for training signals. The metrics (perplexity, entropy, repetition) measure model confidence, NOT correctness. They are useful for:
- Filtering low-quality outputs before returning to users
- Ranking multiple candidate outputs
- Triggering regeneration for concerning outputs
They should NOT be used as training rewards because:
- Confidence != correctness (a model can be confident and wrong)
- Training on confidence creates circular reinforcement
- This leads to mode collapse, not improvement
For research-valid training, see the ttt module (Test-Time Training).
Structs§
- Filter
Result - Result of quality filter evaluation
- Quality
Filter - Quality filter for generation outputs
- Quality
Filter Config - Configuration for quality filtering
Enums§
- Filter
Reason - Reason for filter failure