Update README.md

README.md

@@ -1,5 +1,168 @@
 ---
+pretty_name: Processed HAR Benchmark Collection
 license: other
-license_name: mixed-source-public-datasets
-license_link: LICENSE
+task_categories:
+- time-series-classification
+tags:
+- human-activity-recognition
+- har
+- wearable-sensors
+- smartphone-sensing
+- imu
+- accelerometer
+- gyroscope
+- timeseries
+- benchmark
+- tabular
+configs:
+- config_name: uci_har
+  data_files:
+  - split: train
+    path: data/uci_har/train.parquet
+  - split: test
+    path: data/uci_har/test.parquet
+- config_name: unimib_shar
+  data_files:
+  - split: train
+    path: data/unimib_shar/train.parquet
+  - split: test
+    path: data/unimib_shar/test.parquet
+- config_name: usc_had
+  data_files:
+  - split: train
+    path: data/usc_had/train.parquet
+  - split: test
+    path: data/usc_had/test.parquet
+- config_name: flaap
+  data_files:
+  - split: train
+    path: data/flaap/train.parquet
+  - split: test
+    path: data/flaap/test.parquet
+- config_name: hapt
+  data_files:
+  - split: train
+    path: data/hapt/train.parquet
+  - split: test
+    path: data/hapt/test.parquet
+- config_name: mhealth
+  data_files:
+  - split: train
+    path: data/mhealth/train.parquet
+  - split: test
+    path: data/mhealth/test.parquet
+- config_name: dsads
+  data_files:
+  - split: train
+    path: data/dsads/train.parquet
+  - split: test
+    path: data/dsads/test.parquet
+- config_name: pamap2
+  data_files:
+  - split: train
+    path: data/pamap2/train.parquet
+  - split: test
+    path: data/pamap2/test.parquet
 ---
+
+# Processed HAR Benchmark Collection
+
+This dataset repository provides model-ready, fixed-window versions of eight public Human Activity Recognition (HAR) benchmarks: UCI-HAR, UniMiB-SHAR, USC-HAD, FLAAP, HAPT, mHealth, DSADS, and PAMAP2.
+
+The goal is to make HAR model comparison easier by releasing a consistent set of preprocessed windows, split files, labels, and metadata. The collection covers smartphone-based sensing, wearable IMU sensing, daily activities, postural transitions, fall-related motions, and sports-style activities under different sensor layouts and label granularities.
+
+## Dataset Subsets
+
+| Config | Dataset | Sensing Setting | Classes | Channels | Timesteps | Sampling Rate |
+| --- | --- | --- | ---: | ---: | ---: | --- |
+| `uci_har` | UCI-HAR | Waist-mounted smartphone accelerometer and gyroscope | 6 | 9 | 128 | 50 Hz |
+| `unimib_shar` | UniMiB-SHAR | Trouser-pocket smartphone accelerometer | 17 | 3 | 151 | 50 Hz |
+| `usc_had` | USC-HAD | MotionNode at the front right hip | 12 | 6 | 512 | 100 Hz |
+| `flaap` | FLAAP | Waist-mounted smartphone accelerometer and gyroscope | 10 | 6 | 100 | 100 Hz |
+| `hapt` | HAPT | Waist-mounted smartphone accelerometer and gyroscope | 12 | 6 | 128 | 50 Hz |
+| `mhealth` | mHealth | Body-worn sensors on chest, right wrist, and left ankle | 12 | 12 | 128 | 50 Hz |
+| `dsads` | DSADS | Five wearable inertial units on torso, arms, and legs | 19 | 45 | 125 | 25 Hz |
+| `pamap2` | PAMAP2 | IMUs on hand, chest, and ankle plus heart-rate sensing | 12 | 40 | 171 | 100 Hz |
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("shenjianmozhu/processed-har-benchmark-collection", "uci_har")
+train = ds["train"]
+example = train[0]
+
+signal = example["signal"]
+channels = example["channels"]
+timesteps = example["timesteps"]
+label = example["label"]
+```
+
+The `signal` field stores a flattened window. Reshape it according to `layout`:
+
+```python
+import numpy as np
+
+def restore_window(row):
+    x = np.asarray(row["signal"], dtype="float32")
+    if row["layout"] == "channels_first":
+        return x.reshape(row["channels"], row["timesteps"])
+    if row["layout"] == "timesteps_first":
+        return x.reshape(row["timesteps"], row["channels"])
+    raise ValueError(f"Unknown layout: {row['layout']}")
+```
+
+If you prefer NumPy archives:
+
+```python
+from huggingface_hub import hf_hub_download
+import numpy as np
+
+path = hf_hub_download(
+    repo_id="shenjianmozhu/processed-har-benchmark-collection",
+    repo_type="dataset",
+    filename="arrays/uci_har.npz",
+)
+arr = np.load(path)
+print(arr.files)
+```
+
+## Data Fields
+
+- `id`: stable sample identifier
+- `dataset`: subset slug
+- `split`: split name
+- `signal`: flattened sensor window
+- `channels`: number of sensor channels
+- `timesteps`: number of time steps
+- `layout`: `channels_first` or `timesteps_first`
+- `label`: integer label
+- `label_name`: human-readable label, when available
+- `subject`: subject identifier, when redistribution terms permit release
+
+## Licensing
+
+This repository contains processed versions of public datasets. The original datasets retain their own licenses, citation requirements, and redistribution terms. The `license: other` metadata is intentional because the collection is license-mixed. Users must comply with the terms of each original dataset.
+
+If any source dataset does not permit redistribution of derived/preprocessed files, the corresponding processed files should be removed from this Hugging Face repository and replaced by preprocessing scripts plus links to the original source.
+
+## Citation
+
+Please cite this processed collection and the original datasets used in your work.
+
+```bibtex
+@misc{har_processed_benchmark_collection_2026,
+  title        = {Processed HAR Benchmark Collection},
+  author       = {Qi Teng and collaborators},
+  year         = {2026},
+  howpublished = {\url{https://huggingface.co/datasets/shenjianmozhu/processed-har-benchmark-collection}},
+  note         = {Version v0.1.0}
+}
+```
+
+Original dataset citation keys: `anguita2013public`, `micucci2017unimib`, `zhang2012usc`, `kumar2022flaap`, `reyes2016transition`, `banos2014mhealth`, `altun2010comparative`, and `reiss2012creating`.
+
+## Responsible Use
+
+The data are intended for research on HAR, time-series classification, wearable sensing, and benchmark reproducibility. Users should avoid using the data for identity inference, health-status inference, or surveillance applications beyond the scope of the original datasets and consent protocols.