Fleuret: v1.4 training was a disaster, but FCA-style models hit 41.17% on GSM8K
francoisfleuret · x · 2026-11-15
François Fleuret reports that his v1.4 training run was a disaster—loss spikes everywhere, bad pretrained and SFT cross-entropy. A comparison table shows 28x/36x FCA models (1.5–1.9B params) vs v1.1/v1.4 (1.51B): v1.4 ended at 2.13 pretrained CE and only 33.28% GSM8K accuracy, while v1.1 topped out at 41.17%. He concludes he may have found "a nice class of models for reasoning" in the FCA family.
Related event: Fleuret's Small Model v1.4 Training Falters as FCA Models Top GSM8K(2 posts)→
More from Research
- Stanford's Flash ALOHA robot arms catch a thrown ball in 1.85s with just 50 demos — k7agar · 2026-11-15
- Drone swarms tie knots in midair to lift objects autonomously — lukas_m_ziegler · 2026-11-15
- SO-Arena Benchmark: Top-Scoring Model Ranks Only 7th in Decision Reliability — kalyan_kpl · 2026-11-15
- World Model Hallucination Paper Wins Best Paper at CoRL 2026 PhysWM Workshop — LerrelPinto · 2026-11-15
- RNN memory is set by recurrent matrix eigenvalues — LRU's ring init fixes it — RobertTLange · 2026-11-15
- Ropedia Xperience 2: aligning scene scans and human actions for robot simulation reuse — liuziwei7 · 2026-11-15