Standard
Core task distribution
DATA → BENCHMARK → DECISION
Data and evaluation
Leakage, formatting, and safety checks are designed before looking at loss.
Paraphrase
Surface variation
Missing info
Uncertainty behavior
Negative
Resistance to false premises
Escalation
Safe redirection
Dataset gate
- Train/validation/test sources are disjoint
- Duplicate and near-duplicate scan
- PII and private operational details removed
- Chat template inspected with the real tokenizer
Model acceptance gate
- Domain ≥ baseline
- Format ≥ 95
- Safety ≥ baseline
- Retention drop ≤ 3
- Same seed and generation settings
1 / 20
Quiz progress5%