-
-
Notifications
You must be signed in to change notification settings - Fork 0
Pareto Frontier
Lodri Péter edited this page Jun 25, 2026
·
1 revision
The must-keep loss weight λ controls the precision–compression tradeoff. v8's training data held constant; only λ varies.
| λ | Model | Heretic exact | Compression | keep_rate | Status |
|---|---|---|---|---|---|
| 3.0 | v8 | 0.955 | 15.0% | 0.854 | Production (sweet spot) |
| 5.0 | v17 | 0.963 | 3.7% | 0.963 | Pareto middle |
| 10.0 | v16 | 0.972 | 2.8% | 0.972 | Best precision, near-zero compression |
- Each +1.0 in λ buys ~0.01 heretic precision at ~10% compression cost
- At λ=10, the model keeps 97.2% of tokens — useless as a compressor despite the best precision
- λ=3 (v8) is the knee of the curve: 0.955 heretic with 15% compression
In headroom (the production compression infrastructure), compression aggressiveness matters for context-window savings. λ=3 gives the best tradeoff: meaningful compression (15%) with high must-keep survival (0.955 heretic, 1.000 agent mk_in_ref with override).