AirFM-DDA Dataset

Dataset Summary

This repository contains the precomputed CSI evaluation set used by the AirFM-DDA validation pipeline.

The data is derived from the processed DeepMIMO test split and stores binary PyTorch sample files under samples/, together with a root-level manifest.json that records sample ordering, split metadata, and per-sample paths.

This repository is intended to be used together with the model checkpoints in:

• https://huggingface.co/SII-kejia/AirFM-DDA-Model

What Is Included

Current repository contents:

• samples/: precomputed CSI samples stored as .pt files
• manifest.json: dataset manifest used by the released validation loader
• .gitattributes: Hugging Face LFS tracking metadata
• README.md: this dataset card

Dataset scale:

• Total saved samples: 47,256
• Approximate on-disk size: 122 GB
• Save dtype: float32
• Maximum temporal length (T): 80
• Maximum subcarrier coverage (K): 128

Per-folder counts:

Source and Generation Settings

The current release was generated with the following recorded settings from manifest.json:

• Source multipath root: processed DeepMIMO test data
• Frame configuration file: frame_structure_configs_test2.yaml
• Validation ratio: 0.1
• Split seed: 42
• CSI generation seed: 59
• Number of paths used in CSI generation: 25
• Storage layout: samples/<city_folder>/<cfg_name>/csi_sample_XXXXXXXX.pt

This repository stores precomputed evaluation artifacts, not the original raw DeepMIMO source files.

Directory Layout

AirFM-DDA-dataset/
├── README.md
├── manifest.json
└── samples/
    ├── city_18_denver_3p5/
    │   ├── cfg1/
    │   └── cfg2/
    ├── city_19_oklahoma_3p5/
    │   ├── cfg1/
    │   └── cfg2/
    ├── city_23_beijing_3p5/
    │   ├── cfg1/
    │   └── cfg2/
    └── city_27_rio_de_janeiro_3p5/
        ├── cfg1/
        └── cfg2/

Sample Format

Each .pt file stores one serialized PyTorch dictionary with the following keys:

• CSI_sample: saved CSI tensor
• mask_TK: boolean mask tensor
• Rx_ant_ind: receive-antenna index tensor
• cfg_tensor: frame/configuration tensor
• meta: per-sample metadata dictionary

The released validation code expects:

• CSI_sample to have shape [2, T, K, S]
• manifest.json to exist at the dataset root
• manifest["samples"][i]["relative_path"] to point to the corresponding .pt file

The manifest additionally records, for every sample:

• sample_index
• relative_path
• source_val_index
• source_batch_index
• source_in_batch
• shape_key
• city_folder
• cfg_name
• row_index
• cfg_index

Recommended Usage

Option 1: Download with Hugging Face CLI

hf download SII-kejia/AirFM-DDA-dataset   --repo-type dataset   --local-dir ./AirFM-DDA-dataset

After downloading, the local directory should contain both:

• ./AirFM-DDA-dataset/manifest.json
• ./AirFM-DDA-dataset/samples/

Option 2: Download from Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="SII-kejia/AirFM-DDA-dataset",
    repo_type="dataset",
    local_dir="./AirFM-DDA-dataset",
)

Option 3: Load One Sample with PyTorch

import json
from pathlib import Path
import torch

root = Path("./AirFM-DDA-dataset")
manifest = json.loads((root / "manifest.json").read_text(encoding="utf-8"))
first_rel_path = manifest["samples"][0]["relative_path"]
sample = torch.load(root / first_rel_path, map_location="cpu", weights_only=False)

print(sample.keys())
print(sample["CSI_sample"].shape)
print(sample["mask_TK"].dtype)

Using with the Released Validation Pipeline

The AirFM-DDA validation pipeline in the released codebase expects a dataset root containing manifest.json and samples/ exactly as provided here.

A matching loader looks up the root manifest and then loads individual files using relative_path. In other words, point the validation code to the dataset root, not directly to samples/.

Notes and Limitations

• This is a binary artifact dataset composed of .pt files, so the Hugging Face dataset viewer is not expected to provide an interactive table preview.
• This release is designed for AirFM-DDA evaluation and reproducible validation, not as a raw-source DeepMIMO redistribution.
• The repository currently contains the precomputed evaluation split used by the AirFM-DDA workflow; it is not presented as a full train/val/test benchmark package.