d24 SFT - v2 base + OLMo-3 midtrain (0.757B)

nanochat-style depth-24 decoder (24L x 1536 hidden x 12 heads, SwiGLU / RoPE / RMSNorm, tied embeddings, GPT-2 BPE vocab 50304, 0.757B params, 2048 ctx, LlamaForCausalLM).

Lineage: v2 pretrain (13.1B ClimbMix, WSD) -> OLMo-3 Dolmino-style midtrain (2.245B tok, 1 epoch, sfanm/d24-midtrain-olmo3) -> SFT (nanochat chat mix, 1 epoch, sfanm/d24-sft-mixture).

Eval (greedy, full sets): GSM8K 3.79% | MMLU 32.8% | ARC 35.6%. The OLMo-3 midtrain trades math for general knowledge: MMLU/ARC are notably stronger than the math-heavy d24 variants (~27%), while GSM8K is lower than the reasoning-midtrain SFT (9.86%).

Chat format: ChatML (<|im_start|>role\n...<|im_end|>); see chat_template.jinja.