AndroidAgent-SFT-v2 / README.md
alpharomercoma's picture
card: mark as superseded (v2 blank-intent regression; successors v3/v4), fix row counts to match stats/stats.json
bcaf50f verified
|
Raw History Blame Contribute Delete
6.8 kB
metadata
license: cc-by-4.0
task_categories:
  - image-text-to-text
tags:
  - android
  - gui-agent
  - screen-agent
  - mobile
  - androidcontrol
  - gui-odyssey
  - amex
  - screenqa
  - superseded
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev_odyssey
        path: data/dev_odyssey-*
      - split: dev_amex
        path: data/dev_amex-*
      - split: dev_screenqa
        path: data/dev_screenqa-*
dataset_info:
  features:
    - name: images
      list: image
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
        - name: reasoning_content
          dtype: string
    - name: source
      dtype: string
    - name: split
      dtype: string
    - name: episode_id
      dtype: string
    - name: step_idx
      dtype: int32
    - name: gt_action
      dtype: string
    - name: is_terminal
      dtype: bool
    - name: orig_w
      dtype: int32
    - name: orig_h
      dtype: int32
    - name: img_sha
      dtype: string
  splits:
    - name: train
      num_bytes: 7650322238
      num_examples: 78982
    - name: dev_odyssey
      num_bytes: 103966755
      num_examples: 1000
    - name: dev_amex
      num_bytes: 49948861
      num_examples: 468
    - name: dev_screenqa
      num_bytes: 47725361
      num_examples: 500
  download_size: 7746047195
  dataset_size: 7851963215

AndroidAgent-SFT-v2

Superseded — kept public for reproducibility; do not use for new training. Models fine-tuned on this mix (LFM3-VL-2B, LoRA) regressed on AndroidWorld (3/116 on our development seed vs 15/116 for the AndroidControl-only v1 model). The cause: 8,739 training steps have a blank intent (GUI-Odyssey instructions rejected by the verb/region check, and AMEX, which has no per-step instructions), rendered as an empty <think></think> before the action. Successors: AndroidAgent-SFT-v3 (the same rows with every intent filled; 13/116 on the same seed) and AndroidAgent-SFT-v4 (v3 plus navigate-then-answer question episodes and an installed-app list in the prompt). Both are private for now.

Screenshot-only Android agent SFT mix (one screenshot per step + text history -> one JSON action), same row format as experimentalmachines/AndroidControl-screen-agent-sft (v1). Built by data/build_v2_mix.py (lfm3-android repo) following the Codex dataset-viability review (docs/DATA_V2_PLAN.md).

split rows content
train 78,982 AndroidControl 55,936 (high-level next actions + 5,998 terminal) · GUI-Odyssey 9,979 · AMEX 7,072 · ScreenQA 5,995
dev_odyssey 1,000 GUI-Odyssey official random_split test episodes (<= 1 row per episode)
dev_amex 468 5% of AMEX episodes (hash-held-out, never in train)
dev_screenqa 500 ScreenQA official validation split

Columns: images (1 JPEG, resized like LFM3-VL/Qwen3.5 processors to a pixel-AREA budget of 880x880 = 774,400 px, multiples of 32), messages (system / user / assistant with content = compact JSON action and reasoning_content = short intent), source, split, episode_id, step_idx, gt_action, is_terminal, orig_w, orig_h. Coordinates: integers 0-1000 of the screenshot. Action space: click, long_press, input_text, scroll(up/down/left/right; direction = where new content comes from), open_app, navigate_back, navigate_home, keyboard_enter, wait, answer, status(complete|infeasible).

Filters (revision 3, after the Codex data audit results/codex/codex_data_v2_audit.md and re-verification codex_data_v2_verify.md; counts in stats/stats.json)

  • Executor contract (unchanged AndroidWorld JSON executor): input_text = type + Enter, so a source TYPE immediately followed by PRESS_ENTER or by a soft-keyboard submit tap is ONE input_text (AMEX 477 Enter + 34 taps, GUI-Odyssey 3,103 taps merged; a second submit-like tap ends the episode: 36). scroll = swipe along the screen centre, so source swipes that are short, diagonal or do not cross the central band end the episode (GUI-Odyssey 1,282, AMEX 817). The band test is a heuristic for "the centre-line swipe acts on the same scrollable region", not a proof.
  • Terminals: AndroidControl only (5,998 rows). GUI-Odyssey COMPLETE/INCOMPLETE and AMEX TASK_COMPLETE/IMPOSSIBLE markers proved unreliable on review (pending installs, wrong item starred, information requests ending without an answer) and are quarantined: 0 terminal rows from the new sources.
  • Screen-disjoint dev: train rows whose resized image is byte-identical to any dev image are removed (34); exact duplicate (image + messages) train rows removed (64); verified over every row: 0 train/dev image overlap.
  • Intents: GUI-Odyssey low-level instructions are model-generated; one is kept only if its verb fits the action type and any stated screen region contains the human-recorded point, otherwise blanked (10,132 of 59,084 candidate steps). AMEX has no per-step instruction: blank. Blank intents render as an empty <think></think>.
  • Text: key names typed as text (esc, enter, ...), empty text, and near-miss copies of a quoted string or e-mail address in the goal end the episode (GUI-Odyssey 75, AMEX 148). Paraphrased search queries are kept.
  • Apps: the 14 AndroidWorld third-party app families are excluded by name and package, plus ambiguous aliases (Calendar Pro, voice/sound recorder; any app labelled Calendar in GUI-Odyssey): 180 AC episodes, 307 GUI-Odyssey episodes. Shared built-in apps (Settings, Chrome, Clock, Contacts, Camera, Files) are intentionally kept, as in v1.
  • Caps: <= 3 rows per episode for GUI-Odyssey/AMEX, enforced in every quota pass; AMEX input_text <= 1,000, keyboard_enter <= 500 (13 remain after merging).
  • GUI-Odyssey: random_split train episodes only (dev = its test episodes); phones only (2,743 tablet/Fold episodes dropped); post-hoc context field never used.
  • AMEX: episodes with any need_human_check step (56) or a non-final terminal marker (2) dropped; device/screenshot orientation mismatches quarantined (3).
  • ScreenQA: train split only; <no answer> dropped; AW-app screens/questions dropped; <= 2 questions per screen; 1,000 full-sentence + 5,000 short answers on different questions -> {"action_type":"answer","text":...}; one question with a reviewed incomplete answer removed.
  • AndroidControl: v1 train split rows (official train episodes), high-level mode only, no step hints, portrait only.
  • OCR (Tesseract, fail-closed) is retained in the builder for any terminal row from a new source; none remain.

Attribution

AndroidControl (Google, Apache-2.0); GUI-Odyssey (OpenGVLab, CC BY 4.0); AMEX (Chai et al., CC BY 4.0); ScreenQA (Google Research, CC BY 4.0 annotations) on Rico screenshots (Deka et al.). Source licences and terms apply.