Update dataset card

README.md

@@ -1,31 +1,122 @@
 ---
+license: cc-by-4.0
+pretty_name: HiPredict dataset
+configs:
+- config_name: preview
+  default: true
+  data_files:
+  - split: preview
+    path: preview/*
+size_categories:
+- 100K<n<1M
 tags:
 - hirise
 - mars
+- planetary-science
+- image-reconstruction
 - numpy
 ---
 # HiPredict dataset
 
-Research NumPy patches and data records for HiPredict's HiRISE RED4 reconstruction.
-Code: https://github.com/ardan-S/HiPredict
-Model: https://huggingface.co/ardan-S/HiPredict
+**Paper:** preprint coming soon · **Code:** [ardan-S/HiPredict](https://github.com/ardan-S/HiPredict) · **Model:** [ardan-S/HiPredict](https://huggingface.co/ardan-S/HiPredict)
 
-The corpus contains 586,450 arrays in 51 uncompressed tar shards.
-The arrays retain their original bytes. CSV row order and all non-Path values
-are preserved; paths resolve relative to each downloaded CSV after extraction.
-The released model uses the recorded global normalization statistics supplied
-with its checkpoint and the public code.
+The patches and records used to train and evaluate HiPredict. HiPredict
+reconstructs the RED4 band of the HiRISE camera on the Mars Reconnaissance
+Orbiter. The RED4 detector has partially failed, and many observations have no
+usable RED4 band as a result.
 
-## Download and extract
+## Contents
+
+| Record and arrays | Patch sets | Observations | Arrays | Observation dates |
+| --- | ---: | ---: | ---: | --- |
+| `data_record_alls.csv`, `npy_files/` | 116,790 | 19,465 | 583,950 | 2008-12-21 to 2021-08-08 |
+| `data_record_evals.csv`, `npy_files_evals/` | 500 | 500 | 2,500 | 2007-09-30 to 2023-05-27 |
+
+The arrays are stored in 51 uncompressed tar shards (`shards/0000/part-*.tar`,
+about 216 GB in total). `dataset_manifest.json` lists each shard's SHA-256
+checksum and the provenance of the two records.
+
+Each patch set holds five float32 arrays from one observation:
+
+- BG12, IR10 and RED4: 256 × 256 patches of the same ground;
+- RED3 and RED5: 1024 × 32 × 4 context arrays from the detector edges next to RED4.
+
+Values are in the calibrated units of the processed ISIS cubes.
+
+## Preview
+
+The dataset viewer shows `preview/`: false-colour composites of 100 patch sets,
+spread evenly over the record's dates, with IR10, RED4 and BG12 as red, green
+and blue, each contrast-stretched independently. They are for browsing only;
+the data are the float32 arrays in the shards.
+
+## Records
+
+Each record has one row per array. The columns are:
+
+- `Observation`, `Set`, `Patch` and `CCD`: which observation, patch set and detector the array belongs to;
+- `Path`: the array's location, relative to the CSV file;
+- `Height`, `Width` and `Channels`: the array's shape;
+- `Patch_x` and `Patch_y`: where the patch sits in the processed cube;
+- `Date`, `Binning` and `TDI`: observation settings;
+- observation metadata from the HiRISE EDR index, including instrument temperatures, viewing and illumination geometry, orbit number, and solar longitude and distance;
+- `Pix_min`, `Pix_max`, `Obs_center` and `Obs_scale`: pixel range and observation-level centre and scale.
+
+## Splits
+
+There are no split files. The code splits `data_record_alls.csv` by date: patch
+sets are sorted by observation date, and the earliest 70% are used for training,
+the next 10% for validation and the last 20% for testing. On this record that
+gives 81,753, 11,679 and 23,358 sets, with boundaries on 2018-02-10 and
+2018-11-19. The released model was trained when the record held 51 fewer patch
+sets, so its training split had 81,717 sets.
+
+`data_record_evals.csv` is the evaluation set for the per-category metrics
+defined in the code's `hipredict/configs/evals.yaml`. 314 of its 500
+observations also appear in `data_record_alls.csv` with identical patches, 240
+of them in the training split.
+
+## Processing
+
+Observations were selected from the HiRISE archive in the NASA Planetary Data
+System. Each one has all five detectors used here (RED3, RED4, RED5, IR10 and
+BG12), binning 1 or 2 and TDI 64 or 128, with matching binning across the RED
+detectors and across the colour detectors. They were processed with HiPROC
+0.11.0 and ISIS 9.0.0 using the pipeline in the code repository. Patch positions
+were sampled to spread the patch sets over location, orbit, solar longitude and
+binning/TDI.
+
+## Download and use
+
+To use the dataset with the HiPredict code, download it into the repository's
+`files/` folder and unpack the shards there:
 
 ```bash
-hf download ardan-S/HiPredict --repo-type dataset --local-dir HiPredict-data
-find HiPredict-data/shards -name '*.tar' -exec tar -xf {} -C HiPredict-data \;
+uv run hf download ardan-S/HiPredict --repo-type dataset --local-dir files \
+  --exclude README.md --exclude .gitattributes
+for shard in files/shards/*/*.tar; do tar -xf "$shard" -C files; done
+```
+
+Unpacking needs as much space again as the download. See the code's
+[workflows guide](https://github.com/ardan-S/HiPredict/blob/main/docs/workflows.md#the-hipredict-dataset)
+for storing the dataset on another disk and for training and evaluating with it.
+
+The records can also be used directly:
+
+```python
+import numpy as np
+import pandas as pd
+
+record = pd.read_csv("files/data_record_alls.csv")
+row = record.iloc[0]
+array = np.load(f"files/{row.Path}")
 ```
 
-This restores `npy_files/OBSERVATION_ID/...` for training/validation/test and
-`npy_files_evals/OBSERVATION_ID/...` for the separate evaluation corpus.
-The two directories remain separate, including when patch filenames coincide.
-CSV records: `data_record_alls.csv`, `data_record_evals.csv`.
-Patch sets contain BG12, IR10, RED4, and RED3/RED5 context arrays.
-`dataset_manifest.json` lists the shard checksums and record provenance.
+## Licence and credit
+
+The dataset is released under the Creative Commons Attribution 4.0 licence.
+HiRISE data: NASA/JPL/University of Arizona.
+
+## Citation
+
+A citation will be added here when the preprint is released.