Merged subset of haowu89/open_parallel_think_code_full
and haowu89/open_parallel_think_code_cot.
Total samples: 40,000 (20,000 from full + 20,000 from cot)
Sampling strategy: Balanced (all mid-ratio kept + symmetric ratio=0/1 fill)
No token-length constraint applied
