tempura-localhost/tempura-editor-sessions

Tempura Editor Sessions

v0 pilot · professional video-editing sessions, screen and input paired frame to event

_{A real frame from this dataset: a paid professional editor working in Adobe Premiere Pro, captured by the Tempura recorder.}

Tempura pays skilled video editors to record their real working sessions inside approved editor applications, then pairs what was on screen with what the editor did, moment by moment. That pairing does not exist on the public internet, because it only happens while someone does the work. Every shard in this dataset carries its own consent record, recorder build provenance, and per-artifact SHA-256 lineage.

At a glance

Work sessions	20 accepted, human-reviewed shards
Video	131 ordered mp4 segments, 2.26 GiB, 3.7 h measured footage
Interaction events	31,192 sanitized rows across 8 event kinds
Event vocabulary	flagsChanged, keyDown, keyUp, mouseDown, mouseDragged, mouseMoved, mouseUp, scrollWheel
Editor application	Adobe Premiere Pro (single-app pilot)
Contributors	3 pseudonymous UUIDs, never names or emails
Consent contract	version 2026-10-01 attached to every shard
Digest lineage	33 artifacts server-recomputed, 298 recorder-declared under contract

Uses and Out-of-Scope Uses

This dataset is for internal evaluation of editor-session data quality and packaging. INTERNAL REVIEW ONLY: it is not licensed for training or external release. Identity inference, keystroke reconstruction, and redistribution are prohibited.

Dataset Structure

README.md
meta/
  info.json
  sessions.parquet
  segments.parquet
videos/{work_session_id}/{shard_index}/segment-N.mp4
videos/{work_session_id}/{shard_index}/thumbnail.png
videos/metadata.parquet
data/{work_session_id}/{shard_index}/*.parquet
provenance/{work_session_id}/{shard_index}/*.json

work_session_id plus shard_index identifies a shard. segment_index orders media inside a shard. session_time_s joins structured event rows to elapsed session media time when the source provides a media offset, or to elapsed wall-clock time from capture start for the other streams.

Quickstart

from datasets import load_dataset

sessions = load_dataset("tempura-localhost/tempura-editor-sessions", "sessions", split="train")
segments = load_dataset("tempura-localhost/tempura-editor-sessions", "segments", split="train")
events = load_dataset("tempura-localhost/tempura-editor-sessions", "events", split="train")

Download one work session without cloning the full private dataset:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="tempura-localhost/tempura-editor-sessions",
    repo_type="dataset",
    allow_patterns=[
        "videos/<work_session_id>/**",
        "data/<work_session_id>/**",
        "provenance/<work_session_id>/**",
    ],
)

Data Collection

The native macOS recorder captures one approved editor window. Recording starts and pauses automatically with the approved editor lifecycle, validates locally, and uploads immutable artifacts for server validation and human review. This pilot contains one editor application, Adobe Premiere Pro.

Consent and Lineage

consent_contract_version records the consent contract attached to each shard. Per-artifact SHA-256 values retain provenance labels through the session counts sha256_server_recomputed_count and sha256_recorder_declared_count. server_recomputed means the server hashed stored bytes. recorder_declared_contract means the server accepted the recorder-declared digest after contract checks and did not recompute that SHA-256. The original manifest, consent record, and capture summary are copied under provenance/.

Personal and Sensitive Information

The recorder contract does not capture: typed text or key labels, camera, microphone, biometrics, window titles, file paths, clipboard, or OCR. Residual risk remains because editor UI pixels can show project media being edited; that capture is covered by the contributor consent contract.

Known Limitations

This pilot contains sessions where claimed wall-clock duration far exceeds stored video; inspect yield_ratio rather than treating claimed duration as video duration. Observed yield_ratio ranges from 0.00208239 to 1.01386 in this build. Event streams are decimated after 600 seconds and capped. The pilot covers one editor application, Premiere Pro, and a small contributor pool. Missing codec, dimensions, or actual duration means ffprobe was unavailable or could not read the source segment; values are never fabricated.

---

_{Produced by Tempura · the dataset's lineage is the product}