BEDTime: Benchmark for Evaluating Descriptions of Time Series in Multi-Modal Language Models

Dataset Summary BEDTime is a unified benchmark for evaluating large language and vision-language models on the task of describing time series data. While there are few existing datasets specifically created for time series-to-text generation, many datasets with paired time series and natural language descriptions can be reformulated to fit this purpose. BEDTime does exactly that — harmonizing four existing datasets into a cohesive benchmark.

BEDTime contains a total of 10,175 time series with corresponding natural language descriptions drawn from four sources:

1. TRUCE-Stock:

   • Real-world stock data collected via the Google Finance API
   • 5,687 time series, each 12 time steps long
   • 3 human-written crowd-sourced descriptions per series
   • Descriptions (≤ 9 words) focus on trends, volatility, and relative change

2. TRUCE-Synthetic:

   • 1,677 synthetic time series
   • Each constructed to exhibit exactly one of six predefined patterns (e.g., increase-in-beginning)
   • 3 crowd-sourced descriptions per series (2–8 words)
   • Small noise added for realism

3. SUSHI-Tiny:

   • 1,400 synthetic time series
   • Generated with combinations of trends, periodicities, and fluctuation types
   • Descriptions are synthetically generated based on known generation properties

4. TaxoSynth:

   • 1,400 synthetic time series
   • Lengths range from 24 to 150 time steps
   • Annotated using attribute-based taxonomy and template-based text generation
   • Descriptions are 5 to 40 words long

---

Supported Tasks and Leaderboards

• Time Series Captioning (open-generation)
• Time Series Understanding (true/false(recognition), multiple choice(differentiation))

---

Languages

• English

---

Dataset Structure Each entry includes:

• A time series ( x_i )
• One or more paired natural language descriptions ( d_i )