question: did you have success using it in SFT multiturn datasets?

[felipemello1]

as in the title: did you guys try using renderers at all for SFT multiturn datasets, or just for RL? I wonder if there is any reason to not use it for SFT (e.g. too complex/slow for the trade-offs that SFT doesn't require).

[hallerite]

we haven't really experimented much with it, but we did added support for it in prime-rl! It was added in PrimeIntellect-ai/prime-rl#2496, though there is still an issue related to renderer-powered SFT in prime-rl (PrimeIntellect-ai/prime-rl#2492). The underlying machinery to fix this has already been merged to renderers, though.

It's not entirely clear to me what the best way is to take a multi-turn rollout and render it into n training samples for an arbitrary render, but perhaps doing a turn-level scan could work for most. In RL, we get build up the rollout incrementally, so it's quite parsimonious to deal with renderers that mimic chat templates that do not always extend the prefix (e.g Qwen3 which does not preserve thinking across tool-calls).