Converted from rl-research/dr-tulu-sft-data into OpenAI-native multi-turn messages+tools
(aligned to Zephyr271828/openseeker_v1_sft) plus citation support. Kept long_form+short_form
(dropped exact_answer). 9,031 rows split into 8,941 train / 90 val (1% holdout,
stratified by type, seed 42).
Flat single-turn XML -> native multi-turn (assistant tool_calls + tool messages; think -> reasoning_content).
Tool names… See the full description on the dataset page:
https://huggingface.co/datasets/boyanbiji/dr-tulu-long-form-sft-data.