WorkDemo-350

Website | Paper | Code

English | 简体中文

A desktop demonstration dataset containing task descriptions, screen recordings, action traces, raw input logs, and screenshot evidence. This README covers the package layout, reading conventions, provenance, processing, and limitations.

Subset	Tasks / videos	Final evidence images	Action records
MyPCBench-Bounded Action	64	1,549	2,063
OSWorld	286	6,839	9,888
Total	350	8,388	11,951

Package layout

WorkDemo-350/
  README.md
  README_zh-CN.md
  LICENSE.md
  third_party/
    MyPCBench-LICENSE.txt
    MyPCBench-NOTICE.txt
    OSWorld-LICENSE.txt
  MyPCBench_final_video_64/
    PATHS.md
    simplification_report.json
    task_index.jsonl
    <task-id>/
      episode.json
      task.json
      summary.md
      video.mp4
      trace/
        actions.jsonl
        keylog.log
      evidence/
        manifest.json
        image_actions.jsonl
        *.png
  OSWorld_final_video_286/
    PATHS.md
    task_index.jsonl
    manifest.json
    selected_task_ids.json
    <category>/<task-id>/
      episode.json
      task.json
      summary.md
      video.mp4
      trace/
        actions.jsonl
        keylog.log
      evidence/
        manifest.json
        image_actions.jsonl
        *.png

All 64 MyPCBench tasks are flattened: episode files are directly inside each task directory. OSWorld tasks are grouped by application category. The package contains data and documentation only; no reader, maintenance scripts, or checksum file are included.

Reading the data

Start with the task_index.jsonl in either subset. Each nonempty line is one JSON object containing task_id, instruction, episode, task, video, actions, keylog, action_count, and duration_ms. Paths in this index are relative to the corresponding subset directory, not to the package root.

Follow the indexed episode path to read its metadata, task description, and media. Videos are MP4 files, images are PNG files, and .json files are ordinary JSON. Action and image-action .jsonl files contain one JSON object per nonempty line.

Raw keyboard logs use keylog.log: skip blank lines and lines beginning with # before parsing the remaining JSON records. The comment header contains recording metadata and is intentionally preserved; this is not a strict JSONL file.

Reading the packaged data requires neither network access nor access to the original recording machine. URLs, VM paths, and initialization commands in OSWorld task.json files are original task metadata, not instructions that must be executed to read the demonstrations. Historical source directory identifiers in simplification_report.json are provenance records, not required local paths.

Files and fields

File	Contents and conventions
episode.json	Task/session IDs, recording status, duration, counts, processing flags, and file references. References are relative to the episode directory.
task.json	Task ID and instruction. MyPCBench also includes application/category/difficulty fields; OSWorld may include snapshot and initialization configuration. Fields vary by task.
summary.md	Human-readable episode and action summary; descriptions are derived annotations.
video.mp4	Screen recording used as the visual evidence source.
trace/actions.jsonl	Actions with fields such as kind, media_time_ms, sequence_index, caption, coordinates, and action-specific details. Time is in milliseconds relative to the recording.
trace/keylog.log	Raw input events, typically timestamp, message, and window, preceded by comment metadata.
evidence/manifest.json	Final evidence images, timestamps, action associations, and processing metadata. Image paths are relative to evidence/. Use events[].image to enumerate the evidence sequence.
evidence/image_actions.jsonl	Image/action associations.

Coordinates use the space recorded in each object. Do not assume 0–1000 normalized coordinates or pixel-perfect alignment; see the coordinate limitations below. status=completed describes recording status, not evaluator-confirmed task success.

The OSWorld subset's root manifest.json lists packaged tasks and video hashes. Its total_files and total_bytes cover task directories, excluding subset-level indexes and documentation. selected_task_ids.json is the packaged task list. The MyPCBench simplification_report.json preserves historical processing counts; its episode_dir fields have been updated to the flattened layout. Historical source counts cover the scope stated in that report and should not be interpreted as current packaged file counts.

Provenance and selection

• MyPCBench: recordings for 64 existing Bounded Action tasks. The package does not document how these 64 tasks were sampled from the full task pool.
• OSWorld: 286 tasks retained from 321 recorded tasks. The existing selection excluded 35 tasks where GPT-5.5 scored 1 without skills but less than 1 with v6 skills. This is a post hoc subset based on model results, not a random sample or an official fixed 286-task test set. Estimates of skill improvement on this subset are affected by that selection criterion.

The package does not include per-task human quality acceptance records, recorder counts, recorder training procedures, or consent records. It therefore does not claim a documented manual high-quality selection process. Model score tables are not included; the selection rule above documents provenance and is not a runtime dependency for reading the package.

Processing and evidence retention

Episode metadata records adjacent typing and scroll-event merging while retaining raw input logs. For merged typing evidence, the last frame is retained and its annotation is redrawn. The available MyPCBench processing report records the historical simplification statistics.

Only final evidence images are packaged. Pre-merge historical image directories and their image-path fields have been removed; processing metadata such as merged_event_count and source_action_ids is retained where available. These source action IDs may refer to pre-merge events rather than the current simplified trace. Evidence image counts describe this offline dataset, not a model's inference-time screenshot window.

Recording coordinates and highlights

Recordings were captured while accessing remote desktops through TigerVNC on macOS. Mac screen coordinates, VNC window coordinates, remote desktop coordinates, and video pixel coordinates may differ. Retina scaling, VNC scaling, window position, and capture cropping can affect their correspondence. Similar screen and video resolutions alone do not establish coordinate alignment.

trace/keylog.log preserves captured input records. Highlights, timestamp associations, and action descriptions in the action traces and evidence indexes are derived annotations. They have not undergone complete per-task, per-click coordinate calibration and semantic verification. An artifact_px label is not proof of verified pixel alignment. Assess highlight positions and action descriptions separately: a focus change after a click may cause a description to identify the newly focused control rather than the actual click target.

The retained videos, raw input content, and final evidence images have not been modified to correct suspected coordinate or caption errors. Highlights have not been repositioned to match descriptions. Check annotations against the original video, timestamps, and visible cursor positions. Before using coordinates for supervision or replay, verify coordinate spaces and transformation parameters; do not treat the existing highlights as validated click ground truth.

Quality, privacy, and reproducibility

Package preparation included checks of task counts and IDs, principal file references, structured record parsing, action counts, image references, and file hashes. These checks do not establish semantic correctness, successful task completion, or frame-level synchronization. No validation scripts or full checksum inventory are distributed in the current layout.

A text-pattern scan did not detect common API-key or private-key formats, but did find email addresses and password-related text. All 350 raw keyboard logs are retained and may expose individual keystrokes. These may be test data; their presence alone does not establish whether they are real or synthetic. A complete frame-by-frame media review, reconstruction of typed text, and recorder authorization review have not been documented. The package is not claimed to be fully anonymized.

This is an offline demonstration dataset, not a complete runnable benchmark. It contains no VM images, environment credentials, download caches, or complete evaluator; the packaged OSWorld task files omit evaluator fields. Reading demonstrations is self-contained, but rerunning and scoring tasks requires additional environment resources. No train/test split is supplied: manage splits by task ID to avoid overlap.

Licensing and attribution

Except where otherwise noted, WorkDemo-350 is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). See LICENSE.md for the licensing notice and the license terms. When sharing or adapting the licensed material, credit WorkDemo-350 and credited contributors, preserve applicable notices, link to the license and material source, and indicate changes. Third-party materials retain their applicable licenses.

• The MyPCBench source repository includes an MIT license with the copyright line Copyright (c) 2026 Lawrence Jang. Its license and original notice are included in third_party/MyPCBench-LICENSE.txt and third_party/MyPCBench-NOTICE.txt.
• The OSWorld source repository includes Apache License 2.0, reproduced in third_party/OSWorld-LICENSE.txt.

These upstream texts are preserved for attribution. CC BY 4.0 applies only to rights the dataset licensors can grant; it does not relicense third-party content. An asset rights inventory, maintainer contact, formal citation, and release version have not yet been supplied.