20,000 model-generated rollouts from cosmos1030/qwen3-4b-gmp-tr-dwup-s70pct,
a Qwen3-4B model pruned to 70% sparsity via TR-GMP (dwup-only) NTP+KD training.
Prompts: sampled from open-thoughts/OpenThoughts3-1.2M
(first human turn of each conversation), filtered to ≤1024 tokens after chat-template rendering.
Model: the pruned checkpoint above, served with vLLM.
Sampling: max_tokens=2048… See the full description on the dataset page:
https://huggingface.co/datasets/cosmos1030/qwen3-4b-gmp-tr-dwup-s70pct-ot3-rollouts.
