Question about the SFT data recipe in Draft-OPD

[xinyangATK]

Hi Haodi,
I really enjoyed your work on Draft-OPD, and I had a question about the SFT stage that I was hoping you could clarify.

Could you share what data you used for SFT? Specifically:

• Is the SFT data the same as the OPD data?
• Did you use DFlash's original data mixture/ratio, or did you construct your own dataset?
• Would it be possible to release the related recipe (data sources and mixing proportions)?

This would be very helpful for reproducing and building on your results. Thanks a lot for your time, and for sharing the work.

[bingyang-lei]

Thank you very much for your interest in our work and for your thoughtful question!

To clarify, the SFT data is different from the OPD data. For the SFT stage of both the EAGLE3 and DFlash draft models, we used exactly the same recipe as described in the DFlash paper. Specifically:

“we collect a mixture of around 800K samples from NVIDIA Nemotron Post-Training Dataset V2 (Nathawani et al., 2025) and CodeAlpaca (Chaudhary, 2023).”
(Quote from the original DFlash paper)

We did not construct our own dataset or modify the mixing proportions in SFT stage.

Additionally, if you would like to directly use OPD without going through the SFT process, you can simply use the draft model released by DFlash.

Hope this helps, and please feel free to reach out if you have any further questions. Thanks again for your interest!

[xinyangATK]

Thanks for your clarification!