Create RFT on Fireworks directly

[benjibc]

FIREWORKS_API_KEY=$PYROWORKS_DEV_API_KEY FIREWORKS_ACCOUNT_ID=pyroworks-dev FIREWORKS_API_BASE="https://dev.api.fireworks.ai" eval-protocol create rft --base-model accounts/fireworks/models/qwen3-0p6b --dataset-id svgbench-small --output-model svgbench-agent-small-bchen-2 --evaluator accounts/pyroworks-dev/evaluators/test-livesvgbench-test-svg-combined-evaluation1 --max-context-length 65536 --skip-validations --response-candidates-count 4 --batch-size 128000 --chunk-size 50 --epochs 4 --use-temporal-workflow --use-eval-v3 --max-output-tokens 32768 --max-inference-replica-count 4 --learning-rate 0.00003 --lora-rank 16

---

Note

Adds a create rft CLI to launch Reinforcement Fine-tuning jobs on Fireworks with dataset upload/materialization, training/inference/W&B options, and persists evaluator traces for smoother workflows.

• CLI:
  • New command group create with subcommand rft to create Reinforcement Fine-tuning jobs.
  • Supports dataset sources (--dataset-id, --dataset-jsonl, --dataset-builder, --dataset-display-name), training config (--base-model, --warm-start-from, --epochs, --batch-size, --learning-rate, --max-context-length, --lora-rank, --accelerator-count, --region, --output-model, --evaluation-dataset, --[no-]eval-auto-carveout), inference params (--temperature, --top-p, --top-k, --max-tokens, --n, --inference-extra-body), W&B (--wandb-*), and misc (--rft-job-id, --yes, --dry-run).
  • Auto-resolves evaluator id from local traces or single discovered test; prints dashboard links.
• Core helpers (eval_protocol/fireworks_rft.py):
  • Add utilities to detect dataset builders, materialize datasets to JSONL, create/upload datasets, create RFT jobs, manage evaluator traces, build default IDs, and map API host to app host.
• Upload flow:
  • After evaluator creation, persist a local evaluator trace (incl. detected dataset builder) for downstream create rft usage.

^{Written by Cursor Bugbot for commit d6c76c1. This will update automatically on new commits. Configure here.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting

• @codex fix this CI failure
• @codex address that feedback

[cursor]

Bug: Debug Print Shows Incomplete Request Body

The debug print of the request body (lines 224-225) occurs before the body is fully constructed. After printing, the code adds evaluationDataset and outputModel fields (lines 226-231) and filters out None values (line 234). This means the printed body shows an incomplete/incorrect preview that doesn't match what will actually be sent to the API. The print statement should be moved after line 234 to show the final request body.