OpenTSLM aims at building foundational Time-Series Language Models (TSLMs), a new family of AI models that extend Large Language Models (LLMs) to understand and reason about time-series data (e.g., vitals, prices, telemetry, grid loads, clickstreams, machine logs)
Unlike conventional time-series ML models, which require task-specific architectures, TSLMs integrate time-series data as a native input to LLMs, enabling them to interpret temporal patterns and generate insights in natural language. TSLMs are highly flexible and follow tasks instructed in natural language, allowing sensors to be freely combined or omitted, making TSLMs more robust to missing or faulty inputs (e.g., missing sensors)
In early benchmarks, TSLMs have outperformed prior benchmarks for text-based reasoning and classification. They can caption, interpret, and summarize raw time series data, and propose actions based on raw data using a specific variant called Time Series Language Action Models (TSLAMs), all in natural language ().