unify flags

[xzrderek]

Note

Renames and unifies RFT CLI flags, supports fully-qualified evaluator/dataset resources, adds new training/inference options, and sends jobId via query param with expanded tests.

• CLI (create rft):
  • Rename flags: --evaluator-id→--evaluator, --dataset-id→--dataset, --max-tokens→--max-output-tokens, --n→--response-candidates-count, --inference-extra-body→--extra-body, --rft-job-id→--job-id.
  • Add flags: --gradient-accumulation-steps, --learning-rate-warmup-steps, --mcp-server.
  • Accept evaluator/dataset as IDs or fully-qualified resources; normalize and extract terminal IDs.
  • Improve dataset resolution: prefer --dataset-jsonl, else infer from data loaders or input_dataset, else run detected dataset builder.
• RFT job body/mapping:
  • Map new/renamed flags to trainingConfig and inferenceParameters; improved extraBody handling (string or JSON-serializable dict/list).
  • Include displayName, mcpServer; build dataset resource with override when provided.
• API call behavior:
  • Move jobId to query param reinforcementFineTuningJobId during job creation.
• Tests:
  • Update tests for renamed flags, resource handling, dataset inference precedence, and jobId-in-query; add end-to-end CLI parsing test.

^{Written by Cursor Bugbot for commit 8ab1c92. This will update automatically on new commits. Configure here.}