ScreenMemEval / README.md
dayoon's picture
Add files using upload-large-folder tool
46772c2 verified
|
Raw History Blame Contribute Delete
1.99 kB
---
license: mit
pretty_name: ScreenMemEval
language:
- en
task_categories:
- visual-document-retrieval
tags:
- screenshots
- screen-recording
- retrieval
size_categories:
- 100K<n<1M
configs:
- config_name: sme-h1
data_files:
- split: test
path: sme-h1/questions.jsonl
- config_name: sme-h2
data_files:
- split: test
path: sme-h2/questions.jsonl
- config_name: sme-agent
data_files:
- split: test
path: sme-agent/questions.jsonl
---
# ScreenMemEval
Questions about what was done on a computer, for retrieving screenshots from its screen history. Each question is answered by one or more of the recordings in the history. The screenshots were taken every 5 s at 1280×720 on three computers: two people's (SME-H1, SME-H2; from PSAI, [anaisleila/computer-use-data-psai](https://huggingface.co/datasets/anaisleila/computer-use-data-psai)) and an agent's (SME-Agent; from MyPCBench, [ljang0/MyPCBench-tasks](https://huggingface.co/datasets/ljang0/MyPCBench-tasks)).
| Config | Recordings | Screenshots | Questions (single + multi) |
|---|---|---|---|
| `sme-h1` | 2,105 | 63,538 | 5,111 (4,265 + 846) |
| `sme-h2` | 960 | 12,761 | 3,507 (3,096 + 411) |
| `sme-agent` | 787 | 55,922 | 1,808 (1,654 + 154) |
Each config has:
- `questions.jsonl`: `qid`, `question`, `kind` (`single`: one recording answers it; `multi`: several do), `type`, and `groups`, lists of recording ids. A group is found when a screenshot of any of its recordings ranks in the top 10.
- `frames.jsonl`: `fid`, `recording`, `file`, `order` (position in the history).
- `frames/shard-*.tar`: the images, at `<recording>/<name>.jpg`.
- `latency_sample.json`: the recordings used to measure speed (first 50 screenshots each) and 5 warm-up screenshots.
Recall@10 ranks every screenshot of the config. A single question counts if one of its groups is found, and a multi question counts the share of its groups found. Code: GitHub.
Screens may show personal information, as in the source datasets.