--- language: - en tags: - eda - kicad - gui-agent - multimodal - schematic --- # AI4EDA: local release candidate AI4EDA contains KiCad GUI observations and action records, plus schematic files and a visual-skill smoke subset. This package preserves source records and is prepared for review; inclusion does not establish that every record is suitable for training or redistribution. ## Contents | Component | Records | Description | | --- | ---: | --- | | `annotations/gui_events.jsonl.gz` | 6,832 | Non-quarantined GUI action events, including human and teacher-augmented sources | | `annotations/precision_events.jsonl.gz` | 2,892 | Events from 50 sessions selected by the existing Precision manifest | | `annotations/precision_schematics.jsonl.gz` | 50 | Session-to-final-schematic mappings; final schematics are not intermediate step states | | `annotations/oracle_schematics.jsonl.gz` | 103 | Instruction-to-schematic records; 98 source records have missing or unresolved files | | `annotations/schematic_inventory.jsonl.gz` | 1,850 | Schematic candidate paths, source group, format, size and SHA-256; includes duplicates, external files, five Eagle files and four unknown/invalid files | | `images/gui-events.tar`, `images/precision-events.tar` | 19,448 images total | Before and after screenshots referenced by GUI event annotations | | `schematics/schematic-candidates.tar` | 1,850 files | Files in the candidate inventory, preserving relative paths | | `subsets/oracle-visual-skill-smoke.tar` | 504 records | Existing 372 train and 132 eval records with screenshots and source/target schematics; not independent of other collections | | `metadata/image_manifest.jsonl.gz` | 19,448 | Image path, archive, size and SHA-256 | | `metadata/non_kicad_windows.jsonl` | 31 | GUI events whose recorded active application is not KiCad, for review | GUI rows contain instruction, session and task identifiers, before/after screenshot paths, event type, original action parameters and source type. Screen and window coordinates are distinct; window coordinates correspond to window screenshots. The event sets are not split into training and test data and may share tasks or content. The historical Oracle subset split is retained without a new generalization claim. ## Reconstruct locally Extract the two image archives and schematic archive into the same directory. Annotation paths beginning `images/` and `原理图候选/` are relative to that directory. The Oracle archive extracts into `原理图与截图配对_Oracle/` and its own annotation paths are relative to that directory. ```bash mkdir -p AI4EDA_data tar -xf images/gui-events.tar -C AI4EDA_data tar -xf images/precision-events.tar -C AI4EDA_data tar -xf schematics/schematic-candidates.tar -C AI4EDA_data tar -xf subsets/oracle-visual-skill-smoke.tar -C AI4EDA_data ``` `SHA256SUMS` verifies the packaged files. `metadata/image_manifest.jsonl.gz` describes the images inside the archives. JSONL annotations are gzip compressed and retain their original fields. ## Limitations and provenance The data combines recorded and augmented GUI trajectories, project-generated schematics, and files derived from or associated with external projects. The schematic inventory includes materials with differing or unverified licenses; no blanket license is asserted. Images, action text, window titles, external-source rights, and recording consent require release review. The 31 non-KiCad-window records are retained for completeness and explicitly indexed. Prior experiment output directories and `sessions_1000_v2` were not recovered. The four unknown/invalid schematic candidates are kept and identified in the inventory; do not treat them as valid circuits. Counts refer to records or file paths, not necessarily independent tasks or circuits.