sglucas/dflash-q3-8b_math-code-stem-chat-evol-code_rt3060_mj8_frac-chat-0-05-stem-0-1_nodedupe_decon13
DFlash Qwen3-8B self-rollouts — tokenized, grouped to 2048 tokens Training data for a Qwen3-8B block-diffusion / speculative-decoding drafter whose verifier is the same Qwen3-8B. Every response was generated by Qwen3-8B itself, so the drafter learns the verifier's own output distribution (the DFlash training-data recipe, arXiv 2602.06036, Section 5). This repository holds the final preprocessed artifact (datasets.save_to_disk output): already tokenized with the Qwen3 chat… See the full description on the dataset page: https://huggingface.co/datasets/sglucas/dflash-q3-8b_math-code-stem-chat-evo
| Author / organization | sglucas |
|---|---|
| License (as declared) | other |
| Downloads | 0 (as reported by Hugging Face) |
| Likes | 0 |
| Created | 22 October 2026 |
| Last updated at source | 22 October 2026 |
Connected in the knowledge graph
Recent changes
- 22 Oct 2026 License: none → other
- 22 Oct 2026 Description: none → DFlash Qwen3-8B self-rollouts — tokenized, grouped to 2048 tokens Training data for a Qwen3-8B block-diffusion / speculative-decoding drafter whose verifier is the same Qwen3-8B. Every response was generated by Qwen3-8B itself, so the draft
- 22 Oct 2026 Topics added speculative-decoding, dflash, diffusion-llm, qwen3, tokenized
Changes detected by Silytics between refreshes of the source.
Related papers
More from sglucas
Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 22 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.