Fleuret's Small Model v1.4 Training Falters as FCA Models Top GSM8K

François Fleuret reported that his small model v1.4, trained with unconventional methods he described as "unnatural," performed poorly with loss spikes and weak pretraining/SFT cross-entropy, while his 28x/36x FCA models reached 41.17% on GSM8K, proving better suited for reasoning.

2026-11-15 ~ 2026-11-15 · 2 related posts

• Fleuret: v1.4 training was a disaster, but FCA-style models hit 41.17% on GSM8K — francoisfleuret · 2026-11-15
• François Fleuret: 'unnatural' training tweaks turned a disastrous run into a promising reasoning model — francoisfleuret · 2026-11-15