anton-hugging/TimeOmni-1-7B

TimeOmni-1: Generalized Time Series Reasoning Model

TimeOmni-1 is presented as the first generalized, unified model specifically designed for time series reasoning. Developed by anton-hugging, this model addresses the common limitation of pretrained LLMs lacking temporal priors by employing a two-stage training pipeline. Initially, it undergoes supervised fine-tuning (SFT) to embed temporal knowledge, followed by reinforcement learning (RL) with task-grounded rewards to enhance robustness and reasoning quality.

Key Capabilities & Features

• Unified Time Series Reasoning: Designed to handle a variety of time series tasks, moving towards a "train-once, use-across-tasks" paradigm.
• Two-Stage Training: Combines SFT for temporal prior injection and RL for robust, task-grounded reasoning.
• Performance: Demonstrates strong performance across multiple time series reasoning benchmarks, including both in-distribution (ID) and out-of-distribution (OOD) scenarios, significantly outperforming other time series language models.
• Base Model: Fine-tuned from Qwen2.5-7B-Instruct, inheriting its general reasoning abilities while specializing in temporal data.

When to Use TimeOmni-1

• Complex Time Series Analysis: Ideal for applications requiring deep understanding and reasoning over temporal data.
• Forecasting and Prediction: Suitable for tasks involving predicting future trends or values based on historical time series.
• Research and Development: Useful for exploring generalized approaches to time series AI and developing new temporal reasoning applications.