Xianglong Tan

你好，我下了魔塔上的 AgentInstruct 数据集，但 conversation 都是空值，请问是数据不开源了嘛？

请问有计划支持GUI Agent吗？

1

希望能支持GUI Agent，让Agent操作电脑办公软件

现在的MLLM真的有视觉Reflection能力吗

3

这个reflection能力我觉得有三点： 1. 能意识到任务是否完成 2. 能意识到需要做撤销/回退 3. 如果是视觉设计的任务，还要能感知到设计出来的东西是否违反人类审美 这三点对于MLLM来说，一个比一个难，我们测试过PPT编辑类任务，对于复杂点的PPT编辑，几乎很少有正确的Reflection

Powerpoint native api support

2

Hi there, When I try to use UFO to interact with powerpoint, I find it perform really bad. The log shows that UFO only use mouse and keyboard to control...

[Rollout timeout] Loss rollout while training

4

The Error traceback: ``` File "/home/tiger/.pyenv/versions/3.11.2/lib/python3.11/site-packages/agentlightning/verl/entrypoint.py", line 152, in run trainer.fit() File "/home/tiger/.pyenv/versions/3.11.2/lib/python3.11/site-packages/agentlightning/verl/trainer.py", line 318, in fit metrics = self._train_step(batch_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tiger/.pyenv/versions/3.11.2/lib/python3.11/site-packages/agentlightning/verl/trainer.py", line 95, in _train_step batch, agent_metrics =...

Is the training on-policy?

3

Each step will conduct rollout, so I guess the training is on-policy

Rollout phase takes lots of time, and I find that only 1 of 8 GPU is working when rollout.

1

Is it possible to accelerate rollout by taking advantage of all the GPU?

Is is possible to run training task on an agent built by Golang?

4

I build an agent by Golang. Is possible to train the LLM that this agent use?