File size: 1,911 Bytes
4c46bbb
 
5d8719e
 
 
 
 
 
 
 
 
 
 
4c46bbb
5d8719e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
---
license: mit
tags:
  - air-quality
  - pm25
  - pollution-forecasting
  - gradio
  - numpy
  - kaggle
  - aisehack
pretty_name: PM2.5 Forecasting Demo Data
size_categories:
  - n<1K
---

# PM2.5 Forecasting Demo Data

This dataset stores the lightweight artifacts used by the Hugging Face Space demo:

https://huggingface.co/spaces/sumit1703/pm25-forecasting

The files are precomputed outputs from the ANRF AISEHack Phase 2 Theme 2 Pollution Forecasting project. The Space uses these artifacts only for visualization. It does not run live model inference, training, or torch at runtime.

## Files

| File | Description |
|---|---|
| `demo_preds.npy` | Precomputed PM2.5 forecast predictions for demo windows. Shape: `(22, 140, 124, 16)` |
| `demo_inputs.npy` | Input PM2.5 history windows used for comparison. Shape: `(22, 10, 140, 124)` |
| `lat_lon.npy` | Latitude/longitude grid used for map axes. |
| `demo_stats.json` | Feature normalization/statistics metadata from the preprocessing pipeline. |
| `sample_indices.npy` | Original sample/window indices used by the demo. Shape: `(22,)` |

## Intended Use

This dataset is intended for the companion Gradio demo only. It allows the app to load saved prediction artifacts from the Hugging Face Hub instead of storing `.npy` files inside the GitHub repository.

## Source Project

- GitHub: https://github.com/sumitjadhav1703/pm25-forecasting-demo
- Live Space: https://huggingface.co/spaces/sumit1703/pm25-forecasting
- Kaggle Competition: https://www.kaggle.com/competitions/anrf-aise-hack-phase-2-theme-2-pollution-forecasting-iitd

## Notes and Limitations

- These files are for demo visualization, not full model training.
- The dataset contains a small subset of demo windows, not the complete competition dataset.
- The predictions are precomputed and static.
- The Space downloads these files at startup using `huggingface_hub`.

## License

MIT