---
license: apache-2.0
task_categories:
- image-text-to-text
---

# MobileGym-4B — Android Lab

One run of the **MobileGUI-Gym MBGym harness** (GenericV3 protocol + hint v2 ScreenMemory,
loop_detect 8) with the MobileGym-4B checkpoint (Qwen3-VL 4B SFT) on Android Lab
(138 tasks, 9 apps), 25 rounds max, parallel 8. The harness was ported to Android-Lab
for this evaluation. Sampling: MobileGUI-Gym run_train.sh validation config
(temperature 0.1, top_p 1.0, max_new_tokens 1024). Judge: gemini-3-flash for query tasks;
rule-based for operation tasks.

| Run | Correct |
|---|---:|
| 1 | 52/138 |

Accuracy: **52/138 = 37.68%**.

- `run_1/config.yaml` — eval config (agent, sampling, docker image, AVD).
- `run_1/scores/` — `results.jsonl` (per-task judge/rule outcome) and `total.jsonl`.
- `run_1/trajectories.tar.gz` — native traces (per-step prompts/screenshots/actions,
  incl. hint_v2 notes and ScreenMemory stats).
- `two_runs/` — combined episodes, protocol, summary, package manifest.