low-altitude-rf-worldmodel
收藏资源简介：
该数据集名为 Low-altitude RF world-model training data，pretty_name 为 Low-altitude RF world-model training episodes，语言为 zh 和 en，任务类别为 other，标签包括 radio-propagation、world-model、sionna、multimodal。它是面向低空射频（RF）世界模型训练的数据集，适用于射频传播建模、世界模型训练和多模态学习等场景。README 推荐下载 compact_uint16 完整 NumPy 训练包。当前 committed_on_huggingface 清单含 2747 条动态 UAV RF 实体样本，历史辅助样本为 0 条；数据集正在增量上传，尚不是全量完成声明。完整下载清单 train.jsonl 为每行提供仓库内路径、字节数、SHA256、城市、物理场景组、版本及训练用途。每个 .tar.gz 是可独立读取的完整样本，包含以下内容：rss_dbm_uint16.npy，RSS 逻辑形状为 [30,24,512,512]，轴依次为时间、高度、y、x，采用 NumPy Fortran 存储布局以压缩；quantization.json，量化步长 0.01 dB，偏移 32768，无效码 0，解码结果单位仍为 dBm，相对现有 centi-dBm 训练数据可精确还原，无额外量化损失，且不截断有效值；states.npz、array_schema.json、heights_m.npy、timestamps_s.npy，保留原训练文件的状态、动作、链路、材料、掩码、属性和字段映射，具体以实际样本字段为准；visual/bs、visual/topdown、visual/elevated 三路视觉数据，每路 30 张 PNG，共 90 张，visual/synchronization.json 含同步时间、相机标定及 UAV 标注；metadata.json、config.json、state_schema.json，以及源样本提供的 actions.json、scene_metadata.json；read_sample.py 和转换校验记录。读取仅需 NumPy，无需 HDF5 插件。30 帧时间通常为 0–29 s，具体以 timestamps_s.npy 为准；高度逐样本采样，以 heights_m.npy 为准。解压后可 mmap 读取存储数组；读取完整解码 RSS 约需 720 MiB 内存，训练时可按帧或片段加载。所有发布样本均为训练集，当前不提供独立测试集；同一物理场景的不同格式、版本或重试不能算独立场景，分组请使用 physical_episode_group。历史辅助样本中的光学 UAV 不意味着原 RF 中存在 UAV 实体；主训练与历史辅助清单分别为 dynamic_rf_train.jsonl 和 legacy_aux_train.jsonl。地图主要为真实建筑轮廓挤出几何，部分高度为估计；视觉为合成渲染，不是真实照片。材料参数包含合成分组与动态变化，不是逐栋实测材料。固定预算单次 RT 样本不做统计置信度验证，具体需读取 qa_protocol、confidence_evaluated 和逐样本 metadata。转换校验只验证存储往返、状态及视觉同步，不声称重新验证 RT 的物理误差。原 float32 HDF5 训练包和已上传文件保留在原 data/ 与根目录 manifests/train.jsonl，属于历史存储版本；它们与 compact_uint16 可能是同一物理样本，不要合并计数。服务器原始 RF 归档继续保留，训练包不含原始重型归档、地图网格或长焦视图。
The dataset is named Low-altitude RF world-model training data, with pretty_name Low-altitude RF world-model training episodes, languages zh and en, task category other, and tags including radio-propagation, world-model, sionna, and multimodal. It is a dataset for low-altitude radio frequency (RF) world-model training, suitable for RF propagation modeling, world-model training, and multimodal learning. The README recommends downloading the compact_uint16 full NumPy training package. The current committed_on_huggingface manifest contains 2,747 dynamic UAV RF entity samples, with 0 legacy auxiliary samples; the dataset is being uploaded incrementally and is not yet a full completion declaration. The full download manifest train.jsonl provides, for each row, the repository path, byte size, SHA256, city, physical scene group, version, and training purpose. Each .tar.gz is an independently readable complete sample and contains: rss_dbm_uint16.npy, whose RSS logical shape is [30,24,512,512], with axes time, height, y, x, using NumPy Fortran storage layout for compression; quantization.json, with quantization step 0.01 dB, offset 32768, invalid code 0, decoded values still in dBm, which can be exactly restored relative to existing centi-dBm training data with no additional quantization loss and no truncation of valid values; states.npz, array_schema.json, heights_m.npy, timestamps_s.npy, which retain the original training file states, actions, links, materials, masks, attributes, and field mappings, subject to actual sample fields; three visual data streams visual/bs, visual/topdown, and visual/elevated, each with 30 PNG images, 90 in total, and visual/synchronization.json containing synchronization time, camera calibration, and UAV annotations; metadata.json, config.json, state_schema.json, as well as actions.json and scene_metadata.json provided by the source samples; read_sample.py and conversion verification records. Reading requires only NumPy and no HDF5 plugin. The 30 frames usually cover 0–29 s, subject to timestamps_s.npy; heights are sampled per sample, subject to heights_m.npy. After extraction, stored arrays can be read via mmap; reading the full decoded RSS requires about 720 MiB of memory, and training can load by frame or segment. All released samples are training samples, and no independent test set is currently provided; different formats, versions, or retries of the same physical scene do not count as independent scenes, so use physical_episode_group for grouping. Optical UAVs in legacy auxiliary samples do not imply that UAV entities exist in the original RF; the main training and legacy auxiliary manifests are dynamic_rf_train.jsonl and legacy_aux_train.jsonl, respectively. Maps are mainly extruded geometry from real building footprints, with some heights estimated; visuals are synthetic renderings, not real photographs. Material parameters include synthetic groupings and dynamic changes, not per-building measured materials. Fixed-budget single RT samples are not subjected to statistical confidence validation; specific details require reading qa_protocol, confidence_evaluated, and per-sample metadata. Conversion verification only validates storage round-trip, state, and visual synchronization, and does not claim to re-verify the physical error of RT. The original float32 HDF5 training package and uploaded files are retained in the original data/ and root manifests/train.jsonl as historical storage versions; they may be the same physical samples as compact_uint16, so do not merge counts. The servers original RF archive continues to be retained, and the training package does not include raw heavy archives, map meshes, or long-focus views.
Low-altitude RF world-model training data 数据集总结
基本信息
| 项目 | 内容 |
|---|---|
| 数据集地址 | https://huggingface.co/datasets/PATIENTHZ/low-altitude-rf-worldmodel |
| pretty_name | Low-altitude RF world-model training episodes |
| 语言 | zh、en |
| 任务类别 | other |
| 标签 | radio-propagation、world-model、sionna、multimodal |
| 推荐下载 | compact_uint16 完整 NumPy 训练包 |
| 当前清单规模 | 5164 条；动态 UAV RF 实体 4052 条，历史辅助 1112 条 |
| 统计范围 | committed_on_huggingface；正在增量上传，尚不是全量完成声明 |
| 完整下载清单 | compact_uint16/manifests/train.jsonl |
完整下载清单每行含仓库内路径、字节数、SHA256、城市、物理场景组、版本及训练用途。
样本内容
每个 .tar.gz 是可独立读取的完整样本，包含：
rss_dbm_uint16.npy：RSS，逻辑形状[30,24,512,512]，轴为时间、高度、y、x；NumPy Fortran 存储布局用于压缩。quantization.json：0.01 dB 步长，偏移 32768，无效码 0。解码结果单位仍为 dBm。相对现有 centi-dBm 训练数据可精确还原，无额外量化损失，不截断有效值。states.npz、array_schema.json、heights_m.npy、timestamps_s.npy：保留原训练文件的状态、动作、链路、材料、掩码、属性和字段映射；以实际样本字段为准。visual/bs/、visual/topdown/、visual/elevated/：每路 30 张 PNG，共 90 张；visual/synchronization.json含同步时间、相机标定及 UAV 标注。metadata.json、config.json、state_schema.json，以及源样本提供的actions.json、scene_metadata.json。read_sample.py和转换校验记录。仅需 NumPy，无需 HDF5 插件。
python from read_sample import load_rss rss_dbm = load_rss("extracted_sample", frames=slice(0, 5))
float32 dBm；无效位置 NaN。解码必须先转有符号整数，避免 uint16 下溢。
读取与解码
- 30 帧时间通常为 0–29 s，具体以
timestamps_s.npy为准。 - 高度逐样本采样，以
heights_m.npy为准。 - 先解压后可 mmap 读取存储数组。
- 读取完整解码 RSS 约需 720 MiB 内存，训练时可按帧/片段加载。
- 解码必须先转有符号整数，避免 uint16 下溢。
训练集与分组
- 所有发布样本为训练集；当前不提供独立测试集。
- 同一物理场景的不同格式、版本或重试不能算独立场景，分组请使用
physical_episode_group。 - 历史辅助样本的光学 UAV 不意味着原 RF 中有 UAV 实体。
- 主训练与历史辅助清单分别为
dynamic_rf_train.jsonl和legacy_aux_train.jsonl。
数据来源与限制
- 地图主要为真实建筑轮廓挤出几何，部分高度为估计。
- 视觉为合成渲染，不是真实照片。
- 材料参数包含合成分组与动态变化，不是逐栋实测材料。
- 固定预算单次 RT 样本不做统计置信度验证，具体读取
qa_protocol、confidence_evaluated和逐样本 metadata。 - 转换校验只验证存储往返、状态及视觉同步，不声称重新验证 RT 的物理误差。
历史存储与归档
- 原 float32 HDF5 训练包和已上传文件保留在原
data/与根目录manifests/train.jsonl，属于历史存储版本。 - 它们与 compact_uint16 可能是同一物理样本，不要合并计数。
- 服务器原始 RF 归档继续保留，训练包不含原始重型归档、地图网格或长焦视图。
