Examples

Production-ready training recipes demonstrating PSRL’s capabilities across different RL paradigms.

---

RLVR

Reinforcement Learning from Verifiable Rewards. Train models on math and reasoning tasks with rule-based reward signals.

Algorithms: PPO, GRPO, DAPO

RLVR: Reinforcement Learning from Verifiable Rewards

Agentic RL

Multi-turn agent training with tool use. Models learn to invoke code interpreters and interact with real software environments.

Recipes: ReTool, SWE-agent

Agentic RL

Generative Reward Model

LLM-derived reward signals for open-ended tasks: LLM-as-a-Judge scoring and On-Policy Distillation. (Under development)

Generative Reward Model

---

Recipe Overview

Recipe	Task Domain	Reward Type	Training Backend	Status
DAPO	Math / Reasoning	Verifiable (boxed answer)	FSDP / Megatron	Ready
PPO	General	Verifiable	FSDP / Megatron	TBD
GRPO	General	Verifiable	FSDP / Megatron	TBD
ReTool	Math + Code Interpreter	Verifiable + tool-call shaping	FSDP / Megatron	Ready
SWE-agent	Software Engineering	Test execution (F2P/P2P)	FSDP / Megatron	Ready
LLM-as-a-Judge	General / Open-ended	Judge LLM score	,	TBD
On-Policy Distillation	General / Open-ended	Teacher LLM token-level supervision	,	TBD