FP-Agent Raw Dataset — Front Sample

A small sample extracted from the front of the FP-Agent project's raw_dataset.json (dataset published by the FP-Agent authors on OSF: osf.io/j6b5p).

What's in here

sample_trials.json contains the first 6 complete trials from the start of the file:

Agent	Trials	eventFrames
Browser Use	3	158
Skyvern	3	311

Each trial (record) is complete and untruncated:

{
  "<Agent>": [
    {
      "fpjs_data": [ ... ],          // FingerprintJS + creep browser fingerprint
      "behavioral_data": [           // array of heartbeat POSTs from the browser
        {
          "req_headers": "...",      // ASN / geo IP metadata
          "req_body": "<escaped JSON string containing: visitorId, sourcePage, eventFrames>"
        },
        ...
      ],
      "source": { "task_name": "Shopping task - 1440x900 - los angeles - None", ... }
    },
    ...
  ]
}

eventFrames format

Every step is logged as a frame [code, selector, ...fields, timestamp]. The trailing timestamp is performance.now() — a browser monotonic clock in milliseconds relative to page load (NOT wall-clock epoch).

code	meaning	example
mm / md / mu	mouse move / down / up	['mm','#from',269,398,16622.69]
kd / ku	key down / up	['kd','#search','k',false,20163.29]
i	input/insert	['i','#search','insertText','k',20171.7]
c	click / change	['c','#from','San Francisco',15874.29]
sc / su / sd	scroll	['sc','document',2,0,3881]
p	paste	['p',sel,clipboardText,ts]

So yes, every step carries a timestamp. It enables inter-event timing (IET), bursts, event rate, and dwell features. It does NOT support cross-session absolute-clock alignment.

Notes

• This is only the frontmost ~6 trials (~2.2 MB) of the original 4.26 GB dataset (~1600 sessions, 8 agent classes: Browser Use, Skyvern, Comet, Manus, Human, ChatGPT Agent, Atlas Agent, Claude).
• Source: FP-Agent project, OSF node j6b5p (view-only), authors' public repo README.