rnow-train-jsonl

To format train.jsonl data for reinforcement learning, each line must be a structured JSON object containing fields like messages, rewards, metadata, variables, and tools. This validation ensures schema consistency across your entire RL or SFT dataset.

Validating rewards and tools names in your train.jsonl dataset requires cross-checking entries against the defined functions in rewards.py and tools.py. This ensures only valid reward mechanisms and tool configurations are included in your training data.

Docker or sandbox configuration in train.jsonl entries is needed when your training environment requires containerized execution. The formatting process surfaces guidance for sandbox or Docker usage, applying docker and docker_env fields to isolate agent training and evaluation tasks.

Required fields for an SFT train.jsonl file include a messages array with role and content, plus a rewards array listing evaluation metrics. Optional fields like metadata, variables, tools, and docker configurations can be added to enrich the training context.

Yes, this train.jsonl formatting approach works for both RL and SFT datasets. It standardizes each line as a JSON object and enforces schema consistency, applying the same validation rules to required fields, rewards names, and tool configurations across both training methods.

Train.jsonl schema validation fails when entries lack required fields, contain improperly structured JSON objects, or reference rewards and tools names that do not exist in your rewards.py and tools.py files. Formatting standardizes each line to enforce this consistency.