Skip to content

Data Contracts

Sudhanshu edited this page Nov 11, 2026 · 1 revision

Data Contracts

Qwen training JSONL

One JSON dictionary per physical line:

{"instruction":"What is Alien: Romulus about?","output":"Alien: Romulus is a science-fiction horror film."}

Required fields: instruction and output, both nonempty strings. The validation helper reports line numbers on malformed JSON or missing values. The example above is illustrative, not a verified film synopsis.

Legacy SFT records

(prompt: str, completion: str) tuples, formatted into Instruct: <system>\n<user>\nOutput: <assistant><eos>.

Preference records

(prompt: str, chosen: str, rejected: str) tuples; chosen and rejected must differ. Converted into TRL dataset columns prompt, chosen, rejected with EOS suffix on replies.

Saved artifacts

The Qwen default output folder holds PEFT adapter parameters and tokenizer/config; legacy paths hold adapter snapshots. No standalone merged full-model output is promised.

Clone this wiki locally