This is a small, deterministic subset of
moondream/synth-math-reasoning-v2
for reproducing a non-neural speculative-decoding analogue on qwen35 math
reasoning traces.
The point of this dataset is not to train a model. It is to make a simple
measurement reproducible: how many future tokens can a deterministic program
predict from only the prompt, prior accepted tokens, and a causal train split?
model == "qwen35"… See the full description on the dataset page:
https://huggingface.co/datasets/vikhyatk/qwen-math-traces.
