gray311/mobilegym-trajectories-autoglm-phone-9b
MobileGym Trajectories — AutoGLM-Phone-9B (L1) Agent rollout trajectories collected on the MobileGym simulated-Android environment with AutoGLM-Phone-9B as the policy. One row per episode (rollout), in the spirit of SWE-Gym-style instance datasets. Each row carries the task spec, the full multimodal interaction (screenshots + model responses + parsed actions), and a deterministic reward from MobileGym's JSON-state judge (no VLM judging — verdicts are exact).… See the full description on the dataset page: https://huggingface.co/datasets/gray311/mobilegym-trajectories-autoglm-phone-9b.
MobileGym Trajectories — AutoGLM-Phone-9B (L1)
Agent rollout trajectories collected on the [MobileGym](https://mobilegym.dev) simulated-Android environment with AutoGLM-Phone-9B as the policy. One row per episode (rollout), in the spirit of SWE-Gym-style instance datasets. Each row carries the task spec, the full multimodal interaction (screenshots + model responses + parsed actions), and a deterministic reward from MobileGym's JSON-state judge (no VLM judging — verdicts are exact).
Provenance
Stats
- 120 episodes, 50 success (42%), 0 errors
- termination: 97
COMPLETE, 17MAX_STEPS, 6REPETITIVE_LOOP - task-level: 22 tasks with within-group reward variance (RL-usable), 6 always-pass, 12 always-fail
This is trainable data: the 22 variance tasks have both positive and negative trajectories sharing an initial state, suitable for DPO (success vs fail pairs) and (offline) GRPO (group-relative advantage). The autoglm_L1_good.txt task split lists the 22 RL-usable tasks.
Schema
from datasets import load_dataset
ds = load_dataset("gray311/mobilegym-trajectories-autoglm-phone-9b")["train"]
ds[0]["images"][0] # PIL screenshot
ds[0]["reward"] # 1.0 / 0.0Downstream: RFT = filter success==True (50 trajectories); DPO = pair success vs fail rows sharing task_id (22 tasks); offline GRPO = group by task_id, use reward variance.