A 0.75B base language model (nanochat "d24" shape: 24 layers × 1536 hidden, 12 heads, gpt2 vocab padded to 50304, RoPE / RMSNorm / SwiGLU). This is a continued-pretraining (mid-training) checkpoint — not instruction-tuned. For a chat model, use the SFT below.

Training

Base: sfanm/d24-pretrain-v2-climbmix-13B (13.1B-token ClimbMix WSD pretrain).
Mid-training data: sfanm/d24-midtrain-olmo3-5b — a 5.3B-token chunked subsample of the OLMo-3 Dolmino mix (long docs split into ≤2048-token windows, so olmOCR-PDF + reasoning-trace components are kept at true OLMo-3 proportions).
Recipe: 5000 iters = 1 epoch over the 5.24B-token train split, global batch 512, micro batch 2, seq 2048, lr 1.5e-4 → 1.5e-5 cosine.
Loss: train 1.620 · val (5b split) 1.348 · test (5b split) 2.899.

Lineage

SFT of this model → sfanm/d24-sft-v2-olmo3-5b (chat; greedy GSM8K 4.62 / MMLU 35.1 / ARC 37.0), which RLOO then lifts to GSM8K 11.07.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("sfanm/d24-midtrain-v2-olmo3-5b")
model = AutoModelForCausalLM.from_pretrained("sfanm/d24-midtrain-v2-olmo3-5b")

This is a base model — prompt it as a text completer, not a chat assistant.