MixtureVitaev2-100BT-cooldown-rstar_coder

MixtureVitaev2-100BT-cooldown-rstar_coder

One arm of a 30-arm cooldown-data ablation on top of Qwen3-1.7B-Base, part of the MixtureVitae v2 scaling-experiment suite. Branches from a shared MixtureVitae-v1 100BT WSD base checkpoint (step 19,378 of the 24,222-step schedule, ~80BT seen, end of the constant-LR phase), then anneals for the final ~20BT (4,844 steps, LR decaying linearly to zero) on the dataset described below. Every arm starts from the identical base and spends the same ~20BT cooldown budget, so eval differences reflect the cooldown data only.

• Architecture: Qwen3-1.7B-Base (28L, hidden 2048, GQA 16q/8kv, head_dim 128, RMSNorm eps 1e-6, RoPE theta 1e6, QK-norm, no biases, tied embeddings)
• Tokenizer: Qwen3 tokenizer (Qwen2 BPE, vocab 151936)
• Cooldown dataset (this arm): rstar_coder — rstar_coder : 12.7
• Upstream source: microsoft/rStar-Coder

Eval results

Open-sci language suite (11-task mean, 0-shot) + reasoning suite (gsm8k 4-shot exact-match strict; ifeval 0-shot inst-level-strict; MATH500 0-shot accuracy; LiveCodeBench 0-shot accuracy_avg). All values are percentages.

Full 30-arm comparison table and methodology: see the MixtureVitae v2 scaling-experiment writeup (LAION).