system_knowledge_base/unsloth_formats.md

UNSLOTH DATASET FORMAT RULES

The dataset creator should support multiple output schemas rather than forcing all data into one universal format.

Supported logical dataset types:

• strategy_knowledge
• market_time_series
• chart_image_caption
• chart_analysis_conversation
• trade_decision
• mixed_multimodal

Preferred instruction dataset fields: instruction, input, output, metadata

Preferred conversational field: conversations: [{"from":"system","value":"..."},{"from":"human","value":"..."},{"from":"gpt","value":"..."}]

For multimodal data, retain an image field or image reference according to the training pipeline selected by the user.

Each row should contain provenance metadata whenever possible. Dataset rows must be valid JSONL and independently parseable.

The creator should not promise that a dataset is automatically suitable for every Unsloth model. The selected base model, tokenizer, chat template, sequence length, and trainer configuration must match the generated schema.