Frame2KG-YC2-V2-6Luna

Private regenerated annotations for the same YouCook2-derived frame rows as lewiswatson/Frame2KG-YC2 at revision ef28931f8fd1b767efc1927a5d982b185791215d. The original dataset is unchanged. Images, video/frame identifiers, category, split membership and row order are preserved exactly. These are frame images and video identifiers; raw video files are not included.

Annotations and columns

Graphs were generated from frame images with gpt-6-luna. Captions were generated separately with gpt-6-luna from the selected graphs, without images.

• video_id, frame_number, category: original metadata (string values retained).
• image: original embedded image bytes and basename, stored in standard HF Parquet.
• graph: structured JSON with nodes and edges. Nodes have string id/label, bbox as four int32 values [xmin,ymin,xmax,ymax], float64 confidence, and string appearance/size attributes. Edges have string source, target, predicate.
• graph_ct: lossless derived CT string. It is not a sequence of model token IDs.
• graph_sentence: regenerated graph-derived caption string.

Boxes use normalized 0–1000 integer coordinates. The graph generation importer converted native image pixels to this scale using per-axis nearest-integer rounding; CT performs no additional rescaling or quantization. Confidence is model-reported and uncalibrated.

Splits and overlap

Split	Rows
training	12,512
validation	3,357
testing	2,849
validation_dev	100
train_calibration	256

There are 18,718 primary rows and 19,074 rows when both subsets are included. validation_dev overlaps validation; train_calibration overlaps training. Subset images, metadata, graphs, captions and CT match their parent rows exactly. Do not treat these subsets as independent held-out data.

CT codec and verification

Codec: frame2kg-ct-v2-xyxy1000-1; SHA-256: 9c5ba331a4368dddfc72291773fee8732cda7a5638a6aadeb2c8260a6c4221cd. ct/graph_ct.py provides encode_graph, decode_graph, and strict validation; ct/test_graph_ct.py contains eight test methods, including empty graphs, Unicode, whitespace, literal markers/backslashes, empty text, float precision and invalid inputs.

The original 14-marker node-first grammar ends with <|end_graph|>. Text whitespace and list ordering are retained. Backslashes are doubled and < is escaped as the literal sequence \u003c, allowing marker text inside values. There are no per-node or per-edge end markers. Unsupported types/fields are rejected, never dropped. Atomic CT model training requires separately adding the 14 markers as non-special single-token vocabulary entries, resizing embeddings, and training/saving the added rows. CT-Textual uses the same serialized strings with an unmodified tokenizer.

All 19,074 saved rows passed exact JSON→CT→JSON equality, strict numeric/schema checks and confidence-bit preservation. All retained original metadata/image values, split/frame/video memberships and ordering match the pinned source. Parquet reloads match all seven local columns. verification.json records checks and ordered hashes; SHA256.json records release file hashes. No rows were dropped to pass verification.

Use

from datasets import load_dataset
ds = load_dataset("lewiswatson/Frame2KG-YC2-V2-6Luna", token=True)

Authenticate normally with an account authorized for this private repository. Use revision="<commit_sha>" for reproducible loading. Use the source dataset's citation when citing Frame2KG; underlying YouCook2/source image rights still apply.

Limitations

Annotations are synthetic and cooking-domain specific. Structural and lossless serialization checks do not establish semantic or bounding-box accuracy. Graph labels, relations and self-edges can be wrong; captions can omit information or overstate uncertain graph descriptions. Empty-graph captions are uninformative, not evidence that no objects exist in an image. A 27-row caption-to-graph review found documented caveats; it was not an image-ground-truth accuracy estimate. Full raw attempts and local provenance are retained separately and are not bundled in this clean release. No additional license grant is asserted here.