recursive-tsfm-data
收藏资源简介:
本数据集是用于Recursive TSFM模型训练的数据,旨在为`ecntu/tiny-recursive-tsfm`项目提供可复现、明确且可审计的采样数据源。数据来源于四个池:gifteval(50万行,来自Salesforce/GiftEvalPretrain)、tsmixup(300万行,来自Chronos的training_corpus_tsmixup_10m)、kernel(60万行,来自Chronos的training_corpus_kernel_synth_1m)和gifteval_trainval(115,639行,来自Salesforce/GiftEval的无泄露历史序列),总规模约421.5万行。所有历史序列均在最早评估区域之前截断,不保留预测标签,多变量目标以独立的单变量历史序列表示,与模型训练约定一致。数据集提供两种标准采样配方:zero_shot配方(混合gifteval、tsmixup、kernel,权重分别为0.60、0.25、0.15)和aligned_20配方(混合上述四个来源,权重分别为0.48、0.20、0.12、0.20)。每个数据源包含一个`series`列,存储变长的float32观测序列。数据集适用于时间序列预测任务,特别是递归时间序列预测模型的训练。数据许可为混合许可,组件来源包括Apache-2.0许可的Salesforce数据和其他许可的Chronos衍生数据,使用前需查阅上游许可声明。
This dataset is designed for training the Recursive TSFM model, aiming to provide a reproducible, clear, and auditable sampling data source for the `ecntu/tiny-recursive-tsfm` project. The data consists of four source pools: gifteval (500,000 rows, from Salesforce/GiftEvalPretrain), tsmixup (3 million rows, from Chronoss training_corpus_tsmixup_10m), kernel (600,000 rows, from Chronoss training_corpus_kernel_synth_1m), and gifteval_trainval (115,639 rows, from Salesforce/GiftEvals non-leakage historical sequences), with a total size of approximately 4.215 million rows. All historical sequences are truncated before the earliest evaluation region, without retaining prediction labels, and multivariate targets are represented as independent univariate historical sequences, consistent with model training conventions. The dataset offers two standard sampling recipes: the zero_shot recipe (mixing gifteval, tsmixup, and kernel with weights of 0.60, 0.25, and 0.15, respectively) and the aligned_20 recipe (mixing the four sources with weights of 0.48, 0.20, 0.12, and 0.20). Each data source includes a `series` column storing variable-length float32 observation sequences. The dataset is suitable for time series forecasting tasks, particularly for training recursive time series forecasting models. The data license is mixed (license: other), with component sources including Apache-2.0 licensed Salesforce data and other licensed Chronos-derived data; users should consult upstream license statements before use.
数据集概述
数据集名称:Recursive TSFM Training Data
数据集地址:https://huggingface.co/datasets/emiliocantuc/recursive-tsfm-data
许可证:其他(组件许可复杂:Salesforce 来源声明 Apache-2.0,Chronos 来源声明 license: other,TSMixup 衍生自真实世界序列,不保证所有组件均为纯合成数据或受 Chronos 代码仓库 Apache-2.0 许可覆盖)
任务类别:时间序列预测
标签:时间序列、预测、合成数据
大小写名称:Recursive TSFM Training Data
数据集组成
该数据集包含四个独立的源池,以 Hugging Face datasets 工件形式冻结,确保采样过程显式且可审计。每个源具有一个 series 列,包含变长 float32 观测序列。多元目标已表示为独立的单变量历史序列,与模型的训练约定一致。
| 源名称 | 行数 | 来源 |
|---|---|---|
gifteval |
500,000 | Salesforce/GiftEvalPretrain |
tsmixup |
3,000,000 | Chronos training_corpus_tsmixup_10m |
kernel |
600,000 | Chronos training_corpus_kernel_synth_1m |
gifteval_trainval |
115,639 | 来自 Salesforce/GiftEval 的无泄漏历史序列(截断至所有 GIFT-Eval 项的最早评估区域之前,不保留预测标签) |
训练配方
物理行未预混合,由 manifest.json 定义两个规范配方:
| 配方名称 | 代码顺序中的源 | 混合权重 |
|---|---|---|
zero_shot |
gifteval, tsmixup, kernel |
0.60, 0.25, 0.15 |
aligned_20 |
gifteval, tsmixup, kernel, gifteval_trainval |
0.48, 0.20, 0.12, 0.20 |
zero_shot配方不使用 GIFT-Eval 评估序列的任何历史数据用于训练,但仍包含单独发布的GiftEvalPretrain语料库。aligned_20配方额外引入了无泄漏的历史序列,权重为 0.20。
加载方式
使用 Hugging Face datasets 和 huggingface_hub 加载。示例代码如下:
python from datasets import load_from_disk from huggingface_hub import snapshot_download
root = snapshot_download( "emiliocantuc/recursive-tsfm-data", repo_type="dataset", ) aligned_histories = load_from_disk(f"{root}/gifteval_trainval")
复现
确切的制备程序包含在 scripts/ 目录下,使用固定内联 uv 环境。运行命令:
bash uv run --script scripts/prep_eval_data.py --context_len 8192 uv run --script scripts/prep_train_data.py
清单文件记录了固定的上游修订版本、行数、数据集指纹、过滤参数以及两个采样配方。报告代码快照为 f1525b35fc70f675c01f3da959d8122739dcdd1c。





