Search results

The training performance of basic MLLM RLVR configuration is slow

Status: Open.

#310 In alibaba/ROLL;

· canghongjian opened on Dec 30, 2025

code_dapo agentic task fails

Status: Open.

#309 In alibaba/ROLL;

· canghongjian opened on Dec 30, 2025

关于 LitePPO 的相关配置

Status: Open.

#308 In alibaba/ROLL;

· annn521 opened on Dec 30, 2025

ROLL 的下个版本会支持 FSDP 嘛？

Status: Open.

#307 In alibaba/ROLL;

· coswvw opened on Dec 27, 2025

🎉ROLL 2025 Year in Review「2025开源小结」：Celebrating Community Growth & Open Source Milestones & Honoring Contributors🚀

Status: Open.

#306 In alibaba/ROLL;

· tt0718 opened on Dec 26, 2025

请问ROLL是否支持GLM-4.7的多机多卡lora的DPO训练

Status: Open.

#305 In alibaba/ROLL;

· AIR-hl opened on Dec 24, 2025

RLVR默认配置训练出现问题，造成不断减少回复长度，直至为0或不回答,崩掉模型效果

Status: Open.

#303 In alibaba/ROLL;

· millioniron opened on Dec 22, 2025

ipro的训练代码中奖励函数是不可微的，训练中梯度不会更新，和论文中描述不一致，请问是哪里有问题呢？

Status: Open.

#300 In alibaba/ROLL;

· zfnice opened on Dec 18, 2025

(MathRuleRewardWorker(reward-math_rule-1) pid=12406) [2025-12-18 08:08:36] [worker.py (147)] [WARNING] [reward-math_rule-1 1 / 8][PID 12406] worker has not strategy

Status: Open.

#299 In alibaba/ROLL;

· BaiMeiyingxue opened on Dec 18, 2025

Is the code of Rollpacker open source

Status: Open.

#296 In alibaba/ROLL;

· pkuchensw opened on Dec 17, 2025

[BUG] SFT validation fails when use_sequence_packing is enabled

Status: Open.

#294 In alibaba/ROLL;

· XChen-Zero opened on Dec 16, 2025

Agentic pipeline data prepare

Status: Open.

#291 In alibaba/ROLL;

· Qiao0124 opened on Dec 12, 2025