These examples provide concrete examples to leverage vime in your own RL workflow. Some examples are just demonstrative, but most of them are verifiable with a concrete performance score.
- coding_agent_rl: End-to-end SWE coding-agent RL — a real coding agent (claude-code / codex) edits code in a per-sample sandbox, and the resulting
git diffis graded against the dataset's test harness. - eval_multi_task: Example for supporting evaluation multiple tasks with different configs.
- fully_async: Demonstrates fully asynchronous rollout generation for higher efficiency.
- geo3k_vlm: Training VLMs on a single-turn reasoning task using GRPO on the GEO3K dataset.
- geo3k_vlm_multi_turn: VLM multi-turn training on Geo3k dataset.
- low_precision: Launch recipes for FP8/INT4 training and inference.
- mem_agent: MemAgent long-context RL — chunk-wise memory update, HotpotQA GRPO training, and RULER-HQA evaluation.
- multi_agent: Example of running multi-agent RL with
vime. - on_policy_distillation: On-policy distillation (OPD) with an external vLLM teacher or a Megatron-loaded teacher.
- dspark: DSpark speculative decoding draft model training — colocate and non-colocate modes for accelerating RL rollouts.
- delta_weight_sync: Non-colocated weight sync that ships only the changed bytes over a shared filesystem (training/inference disaggregation), reloading via the vanilla
update_weights_from_diskpath. - reproducibility: Guide to bitwise experiment reproduction using deterministic modes.
- tau-bench: Multi-turn tool-use agent training in tau-bench environments.
- train_infer_mismatch_helper: Algorithmic methods for rollout correction (e.g., TIS, MIS).