finetuningScenario-driven OpenAI fine-tuning dataset docs

Scenario-driven dataset toolkit

Build validated fine-tuning datasets from reusable assistant scenarios.

finetuning turns scenario definitions into deterministic or provider-backed OpenAI chat fine-tuning JSONL. Start offline, validate every export, then opt into OpenAI or Anthropic runs only when a workflow needs model calls.

Scenarios carry assistant role, tools, personas, goals, and stopping rules. Runtime provider choices stay explicit in config files or command flags, with secrets kept in environment variables.

Start with the quickstartBrowse the docs

What it solves

Fine-tuning examples that are generated, inspected, and validated from one source of truth.

Ad hoc scripts make training data hard to review. finetuning keeps domain behavior in scenario JSON, turns those scenarios into canonical rows, and validates the output shape before generated or translated JSONL reaches a training job.

Offline-first workflows

Generate personas, simulate datasets, and run local-pseudo translation without provider credentials.

Tool-calling dataset support

Export canonical OpenAI chat fine-tuning JSONL in plain-chat, tool-decision, or full-tool-trajectory mode.

Validation before training

Check JSONL parsing, message roles, tool-call arguments, tool-result references, duplicate IDs, and summary counts.

Explicit provider paths

Switch to OpenAI or Anthropic adapters only when you choose a provider, model, and API key environment variable.

Built for

Teams that need auditable assistant examples before they send JSONL to training.

Model assistant domains in scenario JSON instead of hard-coding behavior in scripts.

Generate deterministic sample datasets from a clean checkout before adding provider credentials.

Preserve full tool-call loops when the model needs to learn tool choice, tool results, and final responses.

Translate generated datasets while keeping tool calls, tool definitions, metadata, and validation rules intact.

Dataset model

Scenarios, export modes, validation, and provider config stay separate.

scenario JSON

Defines assistant role, business context, persona source, tool inventory, goals, and stopping rules.

export modes

Choose plain chat, tool decision, or full tool trajectory depending on what the example should teach.

validation

Checks message shape, tool-call integrity, result references, duplicate IDs, and dataset summary counts.

provider config

Keeps OpenAI and Anthropic models, base URLs, and API key env vars outside scenario definitions.

Docs map

Start offline, validate the row shape, then add provider-backed runs deliberately.

The docs cover setup, deterministic generation, full tool trajectories, validation, translation, provider configuration, scenario authoring, examples, architecture, and maintenance.

Start

Quickstart

The default path stays offline and deterministic, so you can verify the toolkit before adding provider credentials.

Dataset Workflows

Full tool trajectories

Full tool trajectories are the default export path for tool-calling examples and preserve the complete assistant-tool-result exchange.

Dataset Workflows

Validation

Validation is the required checkpoint before using generated or translated JSONL for training.

Dataset Workflows

Translation

Translation is experimental, but it preserves tool-calling structure and validates output before writing.

Provider Workflows

Provider config

Provider-backed runs are explicit, config-driven, and keep secrets in environment variables instead of checked-in JSON.

Authoring

Scenario authoring

Scenario files are the public modeling surface for new domains, not hard-coded framework behavior.

Authoring

Examples

The checked-in examples show how to model different domains without changing toolkit code.

Reference

CLI reference

The CLI exposes deterministic defaults first and makes provider-backed paths explicit per command.

Reference

Architecture

The toolkit keeps `src/core` provider-neutral while pushing provider SDKs and runtime config into dedicated adapter layers.

Reference

Log conversion deferred

`convert-logs` is a deliberate boundary, not a partially implemented feature.

Reference

Docs site maintenance

The docs app is a small static Astro site that consumes shared presentation from `zuedocs`.

Use finetuning when scenario-authored behavior, explicit provider boundaries, and validator-backed JSONL matter more than one-off dataset scripts.