94,505 fresh (prompt, response) pairs generated to extend a DFlash speculator's
training data with more code and multilingual coverage. Generated in
non-thinking mode (enable_thinking=false) to match how the downstream
speculator is trained and evaluated.
Built in two batches: an initial 59,506-row batch (50K code + 9.5K
multilingual), followed by a second batch continuing from exactly where the
first left off… See the full description on the dataset page:
https://huggingface.co/datasets/inference-optimization/dflash-code-multilingual-teacher-responses-qwen235b.
