Search results

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

Missing ETP Support in mca_adapter Under Expert Parallelism

Status: Open.

#290 In alibaba/ROLL;

· vincent-hyx opened on Dec 11, 2025

rope_factor args will be ignored if do rope_scaling during the initialization of McaGPTModel

Status: Open.

#287 In alibaba/ROLL;

· Kuangdd01 opened on Dec 8, 2025

🚀 [2025/12/8] Recent Updates Summary for ROLL Project

Status: Open.

#286 In alibaba/ROLL;

· PanAndy opened on Dec 8, 2025

ROLL目前是否有考虑一些VLA的模型的整合？我看verl已经弄了一下openvla-oft模型了

Status: Open.

#285 In alibaba/ROLL;

· millioniron opened on Dec 8, 2025

on-policy下，log-prob和old-logprob数值不一致，有一些崩溃现象

Status: Open.

#284 In alibaba/ROLL;

· millioniron opened on Dec 8, 2025

华为硬件的 feature 支持

Status: Open.

#282 In alibaba/ROLL;

· Sh1k17 opened on Dec 7, 2025