List your company
ResearchFormerly WizWand

General Capability on GPQA Diamond

52AccuracyIn-context distillation
2.0815.042840.96May 14, 2026May 22, 2026May 30, 2026Jun 7, 2026Jun 15, 2026Jun 23, 2026Jul 2, 2026
Updated 4mo ago

Evaluation results

MethodLinks
In-context distillationTrained to=—2026.0752
SFT on positive documentsTrained to=—2026.0749
SFT on negated documentsTrained to=—2026.0748
single-basis Goggles (b = 1)Trained to=7002026.0747
GogglesTrained to=7002026.0746
Goggles (Redwood framing)Trained to=7002026.0746
forward KLTrained to=7002026.0745
spectral clip (τ = 3.35)Trained to=7002026.0745
base model (no SFT)Trained to=—2026.0743
PODSModel=Qwen3-8B, Selection Ratio=50%2026.0537.4
Full DataModel=Qwen3-8B, Selection Ratio=100%2026.0536.4
PODSModel=Qwen3-4B, Selection Ratio=50%2026.0535.9
Full DataModel=Qwen3-4B, Selection Ratio=100%2026.0535.8
Lmax = 1 (no curriculum)Trained to=7002026.074