Add a Dataset Formats reference guide

[adohe]

Problem

Dataset shape is easy to misunderstand. Raw GSM8K (question/answer) needs different formatting per algorithm: RL math uses prompt/solutions; SFT expects messages, prompt+response, or text. "GSM8K works with AReno" does not mean it works for every algorithm.

Scope

Add a Dataset Formats page answering: "What columns must my dataset have for each AReno algorithm?"

• Mental model: raw datasets vs AReno training schemas vs loader functions.
• SFT schemas (prompt+response, messages, text) and their loss behavior.
• RL math schema (prompt+solutions as reward metadata, not an SFT target).
• DPO preference schema (prompt/chosen/rejected).
• Dataset loader function contract and when to use one.
• GSM8K examples for both RL and SFT; state plainly raw GSM8K is not SFT-ready.
• Working CLI examples per shape.

Acceptance

• Short, exact, example-heavy; no marketing language.
• Cites existing examples/math/ files.
• No code or CLI changes.

[shakilathasneem9]

I'd like to take this on. I'll work on a Dataset Formats page covering SFT, RL math, DPO schemas, loader functions, GSM8K examples, and CLI usage examples.

[adohe]

@shakilathasneem9 Sure please go ahead, I will assign this issue to you !