---
license: apache-2.0
task_categories: [image-text-to-text]
tags: [mobile-gui-agent, agent-memory, reinforcement-learning, conact]
---

# MemGUI-3K-Verl

RL prompt corpus of the ICLR 2027 submission *FARPO: Folding-Aware Reward-Decoupled Policy Optimization for Long-Horizon Mobile GUI Agents* (project MemGUI-RL; project page:
https://memgui-rl-anonymous.github.io/). Every record is one annotated state of a MemGUI-3K
trajectory in the ConAct conversation format consumed by the FARPO trainer
(https://github.com/memgui-rl-anonymous/MemGUI-RL):

| file | states | trajectories | use |
| --- | --- | --- | --- |
| `memgui3k_train_verl.json` | 73,807 | 2,661 | RL prompts (positive states only are used; folding-aware sampling controls the span:step ratio) |
| `memgui3k_test_verl.json` | 8,296 | 295 | held-out validation (never filtered; 5,941 positive states) |

Each record has `task_id`, `step_number`, `is_positive`, `bad_step`, `raw_response` (the annotated
ConAct response: `<thinking>`, `<folding>`, `<tool_call>`, `<ui_observation>`, `<action_intent>`),
`conversations` (system prompt, user prompt with the folded history / UI memory / recent step record
and the screenshot, reference assistant turn) and `metadata` (impact / reasonableness labels of the
step, trajectory ids).

Screenshots are **not** duplicated here: image references have the form
`MemGUI-3K/images/<task>_<attempt>_step_<k>.png` and resolve against the MemGUI-3K release
(https://huggingface.co/datasets/memgui-agent-anonymous/MemGUI-3K). Set `MEMGUI_IMAGE_ROOT` to the
directory that contains `MemGUI-3K/` when training.