Base:
u-10bei/structured_data_with_cot_dataset_512_v2
Adds (targeted augmentation):
JSON -> TOML (canonical, parse-validated): 1200
Text -> CSV (fixed header + single row): 600
Mix weights:
{
"BASE_STRUCT": 0.7,
"TOML_TASKS": 0.2,
"CSV_TASKS": 0.1
}
Seed: 3407
VAL_RATIO: 0.05
