d24-midtrain-v2-reasoning-3.7B

v2 reasoning-midtrained BASE LM (pre-SFT).

nanochat-style depth-24 decoder — 24 layers × 1536 hidden × 12 heads, SwiGLU / RoPE / RMSNorm, tied embeddings, GPT-2 BPE vocab (50304), 0.757B params, 2048-token context.

Lineage. v2 pretrain (13.1B ClimbMix, WSD; decay-as-midtrain from iter-10000) → reasoning midtrain 3.7B (~40% OpenThoughts / 30% ClimbMix / 15% FineMath / 8% OpenMath / 7% MetaMath).

Metrics. Base checkpoint (pre-SFT) — evaluate after SFT. Corresponding SFT: d24-sft-v2-reasoning-3.7B (GSM8K 9.86%).

Use (base LM)

This is a base language model (post-midtrain, pre-SFT) — use it for text continuation, not chat. EOS is the GPT-2 <|endoftext|> (50256). For a chat model, use the d24-sft-* checkpoints.

from transformers import AutoModelForCausalLM, AutoTokenizer
mid = "sfanm/d24-midtrain-v2-reasoning-3.7B"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype="bfloat16", device_map="auto")
inputs = tok("The derivative of x**2 is", return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**inputs, max_new_tokens=128)[0], skip_special_tokens=True))

Research checkpoint from a from-scratch nanochat-d24 replication (pretrain → midtrain → SFT → RL) on NERSC Perlmutter. Trained on third-party corpora (ClimbMix, FineMath, OpenMath, MetaMath, OpenThoughts, OLMo-3 Dolmino, SmolTalk, …) — see those datasets' licenses; provided as-is for research.