Datasets:

---

anaisleila

/

computer-use-data-psai

like 12

Hi, thanks for publishing this dataset!

We've been working on preprocessing this data for VLM training and discovered a significant issue with keyboard event timing.

Missing press events: The original DuckTrack recorder captures both press (keydown) and release (keyup) keyboard events, but only release events appear in the dataset. The press events seem to have been filtered out during the conversion to parquet?

Delayed/batched release timestamps: More critically, the keyboard release timestamps are often significantly delayed from when the key was actually pressed. For example, in task cmcc8u6yc00va1p1ydsdu52zy, the user types "www.foxnews.com" which is visually complete in the video by ~7s. However, the release events are spread over 1s–53s:

t= 1.4s release w (roughly correct)
t= 5.0s release w (3s late)
t= 7.2s release w (5s late)
t= 11.4s release . (9s late)
t= 19.0s release f (16s late)
t= 40.3s release o (~37s late)
t= 52.9s release n,e,w,s,.,c,o,m (all at same timestamp, ~46s late)

The batch of 8 simultaneous releases at t=52.9s coincides with a window_focus event, suggesting pynput's keyboard hooks are being suppressed or delayed while typing in Chrome's URL bar, then flushing when focus changes.

Interestingly, keyboard releases in Notepad (later in the same recording) show normal ~0.2-0.5s cadence, suggesting this is Chrome/browser-specific.

Would it be possible to republish the dataset with the original press (keydown) events included? These would have accurate timestamps from time.perf_counter() at callback time and would make the keyboard data usable for frame-level action prediction.
The raw DuckTrack events.jsonl files should contain them as {"action": "press", "name": "w", "time_stamp": ...}.

Thanks!