| --- |
| license: cc-by-nc-4.0 |
| --- |
| |
| # Dream-Cubed Natural Representative Sample |
|
|
| This directory is a deterministic, class-stratified sample of the Dream-Cubed Natural dataset. It is provided for reviewer inspection of data quality; the full dataset remains available at https://huggingface.co/datasets/dream-cubed/DreamCubedNatural. |
|
|
| The sample dataset for human-authored data is available at https://huggingface.co/datasets/dream-cubed/DreamCubedHumanSample |
|
|
| The sample includes raw natural biome chunks and processed train/validation examples. |
|
|
| ## Creation Procedure |
|
|
| - Sampling seed: `67` |
| - Last source mode added: `raw` |
| - Raw target: up to `32` chunks per raw source label |
| - Processed target: up to `128` chunks per label per split |
| - Strategy: deterministic stratified sampling without replacement |
| - Output: compact `.npz` raw samples and restored, unsharded processed `.npy` split samples |
|
|
| Exact source paths, selected source indices, sample counts, and output files are recorded in |
| `sample_manifest.json`. Processed split directories also include `source_indices.npy` and fresh |
| `manifest.json` files. Dataset metadata files copied from nearby public dataset files are stored |
| under `metadata/` when available. |
|
|
| ## Inspection Notes |
|
|
| This sample is intentionally small and should not be used for model training or quantitative |
| evaluation. It is meant only to help reviewers inspect file formats, labels, chunk tensors, and |
| overall data quality without downloading the full multi-GB dataset. |
|
|
