Training Reports — Dual-Process LM

v1 — Baseline Joint Training

GPT-2 Small (124M) · tiny.yaml

50,000 steps · March 1–3, 2026 · $27.62 total (spot)

3/5 Targets Met Complete 4 spot instances 3 reclamations survived

AR PPL

26.9

AUROC

0.854

ECE

0.010

Diff Loss

4.13

S1 Accuracy

28.7%

v2 — λ-Rebalanced Joint Training

GPT-2 Small (124M) · tiny.yaml · λ_diff=2.0

50,000 steps · March 4–7, 2026 · $31.44 total (spot)

3/5 Targets Met Complete 15 spot instances 31 reclamations survived

AR PPL

29.65

AUROC

0.863

ECE

0.009

Diff Loss

4.70

S1 Accuracy

22.0%

v3 — λ-Tuned Joint Training

GPT-2 Small (124M) · tiny.yaml · λ_diff=1.3

50,000 steps · March 8–12, 2026 · $40.42 total (spot)

3/5 Targets Met Complete 24 spot instances 15+ reclamations survived

AR PPL

27.99

AUROC

0.870

ECE

0.012

Diff Loss

4.16

S1 Accuracy

26.5%