This dataset is used for the certainty-forcing distillation process in dParallel. We use prompts from publicly available training datasets and let the pretrained model generate its own responses as training data. For LLaDA-8B-Instruct, we sample prompts from the GSM8K, PRM12K training set, and part of the Numina-Math dataset. We generate target trajectories using a semi-autoregressive strategy with a sequence length of 256 and block length of 32. We… See the full description on the dataset page:
https://huggingface.co/datasets/Zigeng/dParallel_Dream_Distill_Data.
