| name | grpo-rl-training |
| description | Plan and orchestrate GRPO training through a native Bun workflow with explicit accelerator ownership. |
| version | 0.3.0 |
| tags | ["post-training","reinforcement-learning","GRPO","reward-modeling","reasoning","structured-output"] |
| dependencies | ["bun"] |
| source | bundled |
| subcommands | ["dry-run"] |
GRPO/RL Training
Use the native Bun workflow to validate GSM8K input, construct a reference GRPO request,
evaluate structured-output rewards, and coordinate a caller-owned training host.
xerxes skill grpo-rl-training --dry-run --dataset ./gsm8k.jsonl
The JSONL file must have string question and answer fields. The final answer is read
from the segment after #### in each GSM8K answer.
The dry run emits a JSON request with the reference Qwen 2.5 1.5B, LoRA, and reward
configuration. It does not allocate a GPU, load a model, run an optimizer, or write model
bytes.
Host-Owned Execution
Actual execution is available only to a Bun/TypeScript embedding application that injects:
- an
accelerator port that starts and waits for a training job;
- a
storage port for run records, metrics, checkpoints, and final-model references;
- optionally, a
reporter for lifecycle events.
Use BundledSkills.GrpoTraining.runBasicGrpoTraining with those explicit ports. The host
chooses its accelerator SDK, remote training service, or JavaScript/WASM backend. The
workflow never uses a Python fallback and never fabricates unavailable accelerator services.
Default Reward Plan
incremental-format: 0.125 for each XML tag and a trailing-text penalty;format: 0.5 for complete <reasoning> then <answer> structure;correctness: 2.0 for an exact extracted answer match.
See the source API under src/skills/grpoTraining/ for the request types, data adapters,
reward functions, and host lifecycle contracts.
