Metric	Baseline	This Model	Improvement
ARC-C (norm)	27.73%	30.03%	+2.30pp
ARC-E	54.12%	55.85%	+1.73pp
GSM8K	0.38%	1.67%	+1.29pp
HellaSwag (norm)	42.99%	42.84%	-0.15pp
PIQA (norm)	66.92%	67.08%	+0.16pp

How It Was Made

This model is a SLERP merge (t=0.2) of two specialist models:

KTO model (exp123): Trained with Kahneman-Tversky Optimization — PIQA champion (67.30%)
DPO constant scheduler model (exp146): Trained with DPO sigmoid + constant LR — GSM8K specialist (2.35%)

Both parent models start from exp035 (SFT on ~100K reasoning + non-reasoning data).

Key Insight

Cross-method merging (KTO+DPO) produces better results than same-method merging. The different alignment algorithms create maximally diverse weight patterns, enabling constructive interference during SLERP interpolation.

Full Pipeline

SmolLM2-135M-Instruct
  → SFT (all data ~100K, lr=3e-4, 1ep, cosine)
  → [Branch A] KTO (10K, lr=5e-5, cosine)
  → [Branch B] DPO sigmoid (10K, lr=5e-5, constant scheduler)
  → SLERP merge (t=0.2, 80% KTO / 20% DPO)

Part of 250+ Experiment Study

This model emerged from a systematic study of 250+ training experiments and 50+ merge variants, documented with 106 scientific findings. See the full report at github.com/lldois/tiny_resoning_llm.

lldois/SmolLM2-135M-Reasoning-ARC-Champion

SmolLM2-135M Reasoning — ARC-C Champion (SLERP KTO+DPO Merge)

Benchmark Results (0-shot, lm-eval-harness)

How It Was Made

Key Insight

Full Pipeline

Part of 250+ Experiment Study

lldois/
SmolLM2-135M-Reasoning-ARC-Champion