CODE DATASET

5,233 unique mechanical parts, with isometric images, text instructions, target drawing programs, and reference DXF drawings.

SPLITS
sft_train.jsonl: 3,929 training samples.
sft_val.jsonl: all 256 original validation samples.
sft_val_effective254.jsonl: the existing formal 254-sample validation subset, not additional data.
test.jsonl: all 1,048 original test samples; 1,045 are eligible under the frozen GT-only mask.

ASSETS
Extract images.zip and reference_dxf.zip into this CODE directory. Each ZIP contains its named subdirectory; JSONL image paths then resolve without edits.
images/: 5,233 PNG input images. reference_dxf/: 5,233 existing DXF ground-truth drawings.

METADATA
sample_index.csv gives actual SFT splits, original source splits, categories, historical tiers, asset paths and exclusion flags.
labels.csv is preserved byte-for-byte from the source and labels the combined 4,185 train/validation pool as train.
evaluation_exclusions.json preserves the established validation/test masks; no original samples have been removed.
asset_manifest.jsonl gives byte counts and SHA-256 for every archived asset. SHA256SUMS verifies uploaded files.
dataset_manifest.json records counts, source versions and unchanged source-file hashes.

SCOPE
The original dataset files contain image/text inputs and assistant Python/ezdxf programs as supervision targets, with no construction/training scripts, model weights, credentials, or experiment logs. A separate evaluation toolkit is now provided in evaluation/.
Training/validation target programs use the final compact representation; test targets retain their original representation, matching the existing experiment data.

EVALUATION TOOLKIT
See evaluation/README.txt for setup, reference data and commands. The toolkit scores existing DXF files using the frozen v5.2-rc5k-final metrics, and includes fixed test1045, paper1041 and validation254 populations. It does not execute generated Python or call models.
The dataset manifest and top-level SHA256SUMS describe the data files; the evaluation toolkit has its own PROVENANCE.json and SHA256SUMS. Dataset records and assets are unchanged.