anton-hugging/TimeOmni-1-7B
TimeOmni-1-7B by anton-hugging is a 7 billion parameter generalized, unified model specifically designed for time series reasoning. It injects temporal priors through supervised fine-tuning and refines reasoning robustness using reinforcement learning with task-grounded rewards. This model excels at diverse time series reasoning tasks, demonstrating superior performance compared to other time series language models and reasoning models.
Loading preview...
TimeOmni-1-7B: Generalized Time Series Reasoning Model
TimeOmni-1-7B is a 7 billion parameter model developed by anton-hugging, representing the first generalized and unified model for time series reasoning. It 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. Subsequently, reinforcement learning (RL) with task-grounded rewards is utilized to enhance its robustness and reasoning quality across various time series tasks.
Key Capabilities
- Unified Time Series Reasoning: Designed to handle diverse time series reasoning tasks within a single model.
- Robust Performance: Achieves top-tier performance on multiple time series benchmarks, including both in-distribution (ID) and out-of-distribution (OOD) scenarios for tasks like Task1 (ACC 90.7/97.5 ID, 87.7/98.3 OOD) and Task2 (ACC 69.3/99.8 ID, 64.0/99.8 OOD).
- Improved Temporal Understanding: Incorporates temporal priors through a specialized training methodology, overcoming limitations of general LLMs.
- Scalable Architecture: Part of a model family that includes 4B and 9B parameter versions, demonstrating consistent performance gains with scaling.
Good For
- Applications requiring advanced time series analysis and prediction.
- Developing systems that need to reason about temporal data patterns.
- Researchers and developers looking for a specialized model with strong performance on time-dependent tasks.