language:
- en
tags:
- video
- evaluation
- proactive-assistant
configs:
- config_name: approved
data_files:
- split: test
path: viewer/approved_4100.parquet
dataset_info:
- config_name: approved
features:
- name: sample_id
dtype: string
- name: group
dtype: string
- name: subclass
dtype: string
- name: review_status
dtype: string
- name: video_uri
dtype: string
- name: sample
dtype: json
- name: review
dtype: json
splits:
- name: test
num_examples: 4100
num_bytes: 26190098
download_size: 7067256
dataset_size: 26190098
StreamMind: 4,100 human-approved samples
The default configuration is approved with one test split. It contains the existing 2,600 PS and CT records and the current 1,500 human-selected CM, MC and RUC records.
| Dataset | Subclasses | Samples |
|---|---|---|
| PS | I: 200, P: 200, R: 200 | 600 |
| CT | G: 500, L: 500, M: 500, T: 500 | 2000 |
| CM | G: 121, L: 136, M: 108, T: 135 | 500 |
| MC | G: 112, L: 154, M: 103, T: 131 | 500 |
| RUC | G: 133, L: 133, M: 106, T: 128 | 500 |
Records
The canonical records are in approved_2600.jsonl and approved_cmr_1500.jsonl. Each record has exactly the same fields (names, count, order and nesting) as its source record in LCZZZZ/proactive-eval-reviewed-20261019: PS and CT from upstream revision 342e61aad5a8ee1aa86d29b5a6b2e5f65d1fb897, CM, MC and RUC from upstream revision 5b0c23bced0ef42c94e9f236bd56470dfde006ec. The selection manifests give the source file and the SHA-256 of each original upstream line.
Only text values were revised; every non-text value (timestamps, response windows, preferred times, frame indices, counts, flags), every ID and the number and order of help points and turns are unchanged.
- Plain ASCII text. All strings use ASCII characters only. Dashes and curly quotes were replaced with plain punctuation, math and chemistry symbols with ASCII notation (for example
a^2,sqrt(3),Delta G,Cl2), units and currencies with words or codes (60 degrees,EUR 15), and accented letters with their unaccented forms. On-screen Chinese, Japanese and Korean text is given as a romanization followed by an English translation, for example'Meiri Dati' (Daily Quiz). - Light wording edits. About 3,700 user instructions, instruction switches, task goals and reference responses were lightly reworded to read more like natural speech, for example removing evaluation terms from user requests and replacing repeated sentence templates. The meaning, every requested concern, the intent type (explicit, implicit, revision, composition, interruption), all facts, numbers, names and on-screen labels were kept. Evidence descriptions, key facts and reactive-question answers were not reworded.
Original video archives
Use the selection manifests and the matching entries in media_index/ to locate each video in the original media/ archives. Archives can contain additional videos. Only the 4,100 canonical rows are benchmark samples. The archive verification files provide source sizes and SHA-256 digests.