Jin: Token-level RL enables mixing instruct and base model fine-tuning
Roger Jin · What is an RL environment? w/ Nous Research's Roger Jin · Apr 29, 2025 · at 13:04
This episode carries Roger Jin's own address, with nobody on the show putting questions to them. It still counts as said, and it is kept out of every score on their page.
Roger Jin discusses the technical advantages of designing reinforcement learning environments that interface with trainers at the token level.
0:00 / 0:17exact quote · 17.9s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Another thing you can do, like, with, in, in, like, the token world is, like, the trainer is now, like, agnostic to, like, chat versus instruct model. And what that means is, like, you can do all, like, the cool, like, R-one-zero kind of style, like, experiments. But it also means that you can mix together instruction instruct and base model fine-tuning, like, all at once.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Roger Jin
Jin: RL enables models to surpass expert labelers and develop self-direction
“The model outperforming expert labelers is, is possible. The model learning, like, self-direction is, like, expected. And yeah, we've seen, like, kind of cool emergent behaviors with, like, you know, like, O-one, O-three, R-one, kind of, like, these, like, thi…”
Jin: Supervised Learning Cannot Train Models by Demonstrating Bad Examples
“There's, like, a kind of deficit, which is, like, there's no way to demonstrate what is bad in supervised learning. You kind of just, you have to tell the model very well, like, what is good.”
Jin: Supervised Models Rarely Outperform Expert Human Labelers
“There are, like, kind of surprising results, like weak to strong generalization, but, like, in general, like, the model outperforming the expert labelers would be, like, kind of very surprising.”
Jin: Policy gradient algorithms function as weighted supervised fine-tuning
“If you kind of, like, look at, if you kind of stare at, like, this part it sort of looks like just, like, weighted supervised fine-tuning, right? Like, you have this, like, log of, like, the probability of a token and, like, some, like, weight on it and reinfo…”
Jin: Open source needs a standard to scale RL environments
“And so, like, there just needs to be some kind of, like, standard for open source developers to all, like, work together to, like, build up this, you know, to scale environments up to, like, millions and millions of environments.”
Jin: Fusing inference and scoring natively enables multi-step and multi-agent RL
“Collect trajectories is a fusion of both these. It handles both inference and scoring, and we deliberately chose that because, like, what happens when you try to, like, do, like, multi-turn, or, like, multi-agent with, like, a separate score function? Then thi…”