MixtureVitaev2-100BT-cooldown-nemo_science_math
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):
nemo_science_math — nemo_science_math : 18.7
- Upstream source:
nvidia/Llama-Nemotron-Post-Training-Dataset
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.
| open-sci(11) | gsm8k | ifeval | MATH500 | LCB |
|---|
| 41.9 | 34.3 | 25.8 | 35.4 | 7.8 |
Full 30-arm comparison table and methodology: see the MixtureVitae v2 scaling-experiment writeup (LAION).
