appgen-sft-data / README.md
luca0621's picture
dataset card
c195ae1 verified
|
Raw
History Blame Contribute Delete
1.84 kB

appgen-sft-data — SFT datasets for mobile-GUI agents

Step-level SFT samples harvested from the GRPO simulator environments (appgen-training-data). One sample = one action step:

{
  "messages": [
    {"role": "user", "content": "<image> Goal: <task with app name> History: step1: ...; step2: ..."},
    {"role": "assistant", "content": "<think>one-sentence reasoning</think>\n<tool_call>\n{\"name\": \"mobile_use\", \"arguments\": {...}}\n</tool_call>"}
  ],
  "images": ["<abs path to page screenshot>"],
  "action_type": "TAP|SWIPE|open_app|TYPE|PRESS_ENTER|TASK_COMPLETE|...",
  "env_dir": "...", "task_id": "...", "step": 0, "path_length": 7
}
  • qwen25 variants: absolute coordinates in the smart_resize'd image space (click [x,y]); qwen3 variants: 0-1000 normalized.
  • Entry mix ≈ 54% open_app / 46% drawer-swipe across all versions.
  • Every trajectory ends with terminate(status="success") at the target page.

Current (use these)

file model trajectories notes
sft_qwen25_UNIFIED_V3.json Qwen2.5-VL (abs) 554 traj / 6,853 samples / 100 envs length-uniform: path 3–20 ≈30 each; thinks 100%
sft_qwen3_UNIFIED_V3.json Qwen3-VL (norm) same content, normalized coords

Images: extract unified_sft_assets.tar.gz (preload_v3 envs) and v4sel_sft_assets.tar.gz (50 fresh-pool envs), then rewrite the two path prefixes /data/appgen/preload_v3 and /data/appgen/training_env_pool_v4 to your extraction root. System prompts: unified_prompts.tar.gz.

Older versions (provenance / ablations)

  • sft_qwen25_UNIFIED_V2.json — UNIFIED + easy band only (length-skewed; superseded by V3)
  • sft_qwen25_UNIFIED.json, sft_qwen3_UNIFIED.json — long-task-only originals
  • sft_data_mobile_*_ALIGNED.json — earlier aligned variants