The first open-source successful RL attempt on already long-COT finetuned models of simialr sizes under light budget. Light-R1-14B is also the State-Of-The-Art 14B math model with AIME24 & 25 scores 74.0 & 60.2, outperforming many 32B models.
7b
14b
32b
2,562 Pulls Updated 2 weeks ago
f4d24e9138dd · 148B
{
"stop": [
"<｜begin▁of▁sentence｜>",
"<｜end▁of▁sentence｜>",
"<｜User｜>",
"<｜Assistant｜>"
]
}