Will there be a technical report for V2.5? Curious about SFT stages
#19
by maksimcrewceo - opened
Hi, first of all thank you for your work
Are you planning to release a technical report?
I'm particularly interested in the training stages specifically the SFT pipeline.
It would be great to know:
What was the order of stages (e.g., text-only SFT → multimodal SFT → long-context extension → agentic SFT)?
Did you use curriculum learning (progressively harder/longer tasks)?
How did you balance "thinking" vs "non-thinking" modes across stages?
Any details would be super helpful for local fine-tuning.
Thanks!