Standard
Core task distribution
DATA → BENCHMARK → DECISION
Data and evaluation
Design leakage, formatting and safety checks before inspecting loss. The mix and acceptance thresholds below are teaching examples for the capstone, not universal standards. Scores use a 0–100 scale; retention change is in percentage points.
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
Example model acceptance gate
- Domain ≥ baseline
- Format ≥ 95
- Safety ≥ baseline
- Retention drop ≤ 3 percentage points
- Same seed and generation settings
1 / 20
Quiz progress5%