MM-ArkBench Full Staging
This is the full-staging build of MM-ArkBench: a repository-level multimodal ArkTS / ArkUI dataset for screenshot-to-code retrieval.
Current contents:
- 3,977 repository records from the 4,015 retained ArkTS candidates that could be acquired/parsed.
- 421,107 ArkTS corpus file units.
- 3,274,181 extracted symbols.
- 8,686 image queries.
- Query types:
repo_screenshot,runtime_screenshot, anddoc_or_promo. The current build includes 122runtime_screenshotrows from real build-install-launch-capture runs on a connected HarmonyOS/OpenHarmony device, not from repository assets. - 3,581 repository content families used for split grouping.
The four configs are:
queries: image queries plus OCR/context/provenance and automatic positives where applicable.corpus: retrievable ArkTS file units.symbols: extracted ArkTS/ArkUI components, functions, methods, classes, and builders.repositories: repository provenance, license classification, commits, and acquisition/parse status.
Query type semantics
repo_screenshot: UI screenshots found in repository README/docs/assets, with automatic ArkTS file/symbol candidates.runtime_screenshot: real device/emulator screenshots from build-install-launch-capture, linked back to repository-local ArkTS files.doc_or_promo: documentation, architecture, or promotional images. These rows retain source provenance but intentionally have emptypositive_file_idsandpositive_symbol_ids.
README/docs provenance is stored in matching_evidence, for example:
source:readmesource:docssource:repository_assetreadme:markdown_path=...readme:heading=...readme:line=...
Loading
from datasets import load_dataset
queries = load_dataset("hreyulog/Arkts-mm-ui-full-staging", "queries")
corpus = load_dataset("hreyulog/Arkts-mm-ui-full-staging", "corpus")
symbols = load_dataset("hreyulog/Arkts-mm-ui-full-staging", "symbols")
repositories = load_dataset("hreyulog/Arkts-mm-ui-full-staging", "repositories")
License and redistribution policy
Repository metadata, fixed commits, paths, hashes, and provenance are retained for all selected repositories. Code text and image bytes are included only when the repository/image source is classified as redistributable by the local license policy. For unclear or non-redistributable cases, rows retain reconstruction metadata rather than redistributing protected content.
Labels are automatic and should be treated as noisy until reviewed. The included review_app/ and annotation_review.csv are intended for human verification.
- Downloads last month
- 27