Add littlerock-1M base checkpoint (OpenCerebral)

#74

Adds opencerebral/littlerock-1M, a public 1,002,720-parameter model pretrained from scratch on 5B tokens (40% FineWeb-Edu / 60% DCLM-baseline). No finetuning. Loads with plain transformers (stock LlamaForCausalLM, no trust_remote_code).

This supersedes #72, which submitted a finetuned checkpoint. That one is withdrawn: it scored 35.5% on your private held-out set but 21.5% on the variety set, so its arithmetic does not survive rephrasing. Thanks to @danaxiomic and the team for catching it and for explaining the three-set methodology.

Evaluation: lm-eval 0.4.12, zero-shot, acc_norm, float32, CUDA. ArithMark-3 via bencharithmark-3.py (float32, default acc_norm).

Model ARC-Easy ARC-Challenge HellaSwag PIQA ArithMark-3 Int Index
littlerock-1M 30.98 21.76 26.43 52.23 25.20 2.29

ArithMark-3 is at chance, which is the honest number for a 1M-parameter base model with no arithmetic training.

Architecture: 6 layers, d=96, 6 heads, SwiGLU 224, RoPE, RMSNorm, tied embeddings, 512 context, 4,096-token custom BPE (a stock 50k vocab would cost several times the whole parameter budget at this width). Trained in 1h49m on one RTX 3060.

Datdanboi25 changed pull request status to merged

Sign up or log in to comment