Inquiry About SFT α (alpha) Thresholds

[chaitanyamanem]

Dear Light-R1 Team,

I am studying your Light-R1 model and I’m particularly interested in the two-stage SFT data filtering process described in your paper and repository.

In the paper, the selection of SFT data in Stage-1 and Stage-2 is described using a pass-rate cutoff parameter α, but I could not locate the exact numeric values of α in the paper, the GitHub code.

Could you please share the numeric α thresholds used for Stage-1 and Stage-2 SFT filtering? This would be very helpful for understanding and reproducing your results.

Thank you very much for your time and for releasing this excellent work.

[HaoshengZou]

@juemifuji

[juemifuji]

Dear Light-R1 Team,

I am studying your Light-R1 model and I’m particularly interested in the two-stage SFT data filtering process described in your paper and repository.

In the paper, the selection of SFT data in Stage-1 and Stage-2 is described using a pass-rate cutoff parameter α, but I could not locate the exact numeric values of α in the paper, the GitHub code.

Could you please share the numeric α thresholds used for Stage-1 and Stage-2 SFT filtering? This would be very helpful for understanding and reproducing your results.

Thank you very much for your time and for releasing this excellent work.

"In Phase 1, DeepScaleR-1.5B-Preview was used, with a pass rate of less than 0.2; in Phase 2, deepseek-r1-distill-32B was used with a pass rate of <0.2, and deepseek-r1 sampling 2/3 times had a pass rate of <1."