You.com、NVIDIA 与 CoreWeave 后训练 Nemotron 3.5 Lightning
精选理由
You.com 拉上 NVIDIA 和 CoreWeave 花三周微调 Nemotron 3.5 Lightning，BrowseComp 准确率从 36.97% 提到 45.45%，工具调用还少三成，方法细节都公开了。
You.com 与 NVIDIA、CoreWeave 对 NVIDIA Nemotron 3.5 Lightning 做了后训练，方法是用 SFT 和 RL Rollouts。训练后模型在 BrowseComp 基准上的准确率从 36.97% 提升到 45.45%。同时每个任务的工具调用次数减少了 30%。整个项目在三周内完成。
原文 · CoreWeave
@youdotcom, @nvidia, and CoreWeave post-trained NVIDIA Nemotron 3.5 Lightning using SFT and RL Rollouts.
The results: BrowseComp accuracy rose from 36.97% to 45.45%, with 30% fewer tool calls per task—and we did it in a three-week sprint.
Read all the details and see what you can pull into your own post-training here: https://t.co/TDzFMtQYb8
