Fine-tuning datasets
Dataset format
Write one UTF-8 JSON object per line with state and nonempty questions. Inline questions require type, instructions, and answer; Choice and Score also require criteria.
{"state":"Checkout is unavailable.","questions":{"urgent":{"type":"noul","instructions":"Does this require immediate attention?","answer":1},"team":{"type":"choice","instructions":"Which team owns this?","criteria":{"payments":"Payments","support":"Support"},"answer":"payments"}}}
| Type | Answer |
|---|---|
| Noul | A probability from 0 to 1. |
| Choice | An option key or a probability distribution keyed by option. |
| Score | A zero-based rubric index or a probability array in rubric order. |
Question reuse
Put a question manifest on the first line to reuse its definitions. Later rows can supply only an answer or replace a definition with a complete inline question.
{"urgent":{"type":"noul","instructions":"Does this require immediate attention?"}}
{"state":"Checkout is unavailable.","questions":{"urgent":{"answer":1}}}
Image assets
Rows can define up to ten named PNG, JPEG, or WebP assets (10 MiB each, no animation) via {type: "image", format, data_base64} or {type: "image", url} with public HTTPS. The current mo-1.3.1 fine-tuning base rejects rows that reference images.
Dataset splits
Provide separate validation and held-out files to control the splits. Modaic selects omitted partitions deterministically from the training file and rejects duplicate inputs across files.