TLDR: Datarus-R1-14B is a new language model designed as a virtual data analyst and problem solver. It’s trained on full analytical processes, including reasoning, code execution, and error correction, rather than just question-answer pairs. The model features dual reasoning modes (agentic for tool use, reflection for concise thoughts) and demonstrates an “AHA-moment” pattern, efficiently solving complex problems. It outperforms similar-sized models and rivals larger ones in accuracy on benchmarks like AIME and LiveCodeBench, while using significantly fewer tokens, making it highly efficient and cost-effective.
A new language model named Datarus-R1-14B has been introduced, designed to act as a virtual data analyst and a high-level problem solver. This model, fine-tuned from Qwen 2.5-14B-Instruct, stands out because it learns from complete analytical processes rather than just simple question-answer pairs. This includes every step of reasoning, code execution, error handling, self-correction, and final conclusions, all captured in a notebook-like format.
Datarus-R1 is trained on what the researchers call “analytical trajectories.” These trajectories cover a wide range of quantitative fields, including finance, medicine, and numerical analysis. The training approach is unique, combining a synthetic data generator that produced 144,000 detailed notebook episodes, a special dual-reward system, and an optimized training method called Group Relative Policy Optimization (GRPO).
One of the core ideas behind Datarus is its ability to reason in two ways. In “agentic mode,” it generates steps that involve using Python tools to run actual code, allowing for interactive data analysis. In “reflection mode,” it produces concise summaries of its thought process, similar to a Chain-of-Thought (CoT) approach, using specific tags to structure its output.
The model shows an interesting “AHA-moment” pattern when tackling complex problems. It can form initial ideas, refine them through one or two revisions, and then arrive at a solution, avoiding the repetitive and token-heavy loops often seen in other AI systems. This efficiency is a major advantage. On standard benchmarks like AIME 2024/2025 and LiveCodeBench, Datarus-R1-14B-Preview not only outperforms models of similar size but also competes with much larger reasoning models, achieving up to 30% higher accuracy while using 18–49% fewer tokens per solution. This means it’s both accurate and cost-effective.
The development of Datarus-R1 involved four key innovations. First, the team created a “Trajectory-Centric Synthetic Data Generation” process. They extracted knowledge from various technical sources and used a larger language model, Qwen2.5-72B-Instruct, to generate Python scripts. These scripts created synthetic datasets with specific challenges, which were then executed in a controlled environment. Every step, including thoughts, code, execution results, and errors, was recorded, resulting in 144,000 high-quality problem-solving paths. These paths were categorized to teach optimal solutions, error recovery, self-correction, and how to avoid unproductive approaches.
Second, a “Dual Reward Framework” was implemented. This system uses a tag-based reward to encourage clear and organized outputs, ensuring the model follows a structured format. Additionally, a Hierarchical Reward Model (HRM), built on Qwen2.5-3B, evaluates both individual steps and entire problem-solving sequences. It rewards corrected mistakes and uses preference learning to teach why one reasoning path is better than another.
Third, “Adaptive Curriculum Optimization” was used during training. This involved a gradual shift in focus from structural formatting to semantic correctness. Early in training, the model learned to adhere to the correct output structure. As training progressed, the emphasis moved to the accuracy and quality of the solutions, preventing the model from losing its structured output while improving its analytical skills.
Finally, the “Dual Reasoning Interfaces” provide flexibility. The agentic mode is ideal for interactive analysis, allowing the model to call tools for tasks like data loading or statistical tests. The reflection mode is designed for concise documentation, providing complete reasoning chains in a compact format, useful for reports or proofs.
The training methodology for Datarus involved two main phases: Supervised Fine-Tuning (SFT) followed by Group Relative Policy Optimization (GRPO). The SFT phase established the structured reasoning abilities using the synthetic trajectory dataset. GRPO then refined the model’s performance using the dual reward system, balancing structural formatting with semantic accuracy. The researchers also focused on preventing “overthinking” by identifying and filtering out unproductive repetitions and verbose reasoning patterns in the training data. This ensures the model is concise without sacrificing depth.
The GRPO implementation included significant engineering efforts to optimize memory and computational efficiency, such as KV-Cache reuse, sequential generation processing, and reference model sharding. The tag-based structural reward system specifically encourages the model to start responses with proper structure and use semantic tags appropriately. The Hierarchical Reward Model evaluates the nuanced aspects of data analysis, including process-level correctness and trajectory-level success, even rewarding sequences where errors were successfully corrected.
A dynamic lambda scheduling approach was used to smoothly transition the model’s learning from structural fidelity to semantic depth. This curriculum ensures that the model first learns to format correctly and then refines its solution quality. The distributed generation pipeline, integrated with vLLM, ensures efficient and synchronized generation of multiple solution attempts during training.
The evaluation results confirm Datarus-R1-14B-Preview’s strong performance across various benchmarks, including code generation, mathematical reasoning, and scientific domain knowledge. Its ability to adapt reasoning depth to task complexity, avoiding unnecessary verbosity, is a key highlight. This efficiency translates directly into lower inference costs and faster response times, making Datarus a highly practical solution for real-world applications. The model weights and an interactive agentic pipeline are available for community use, which can be found at the research paper link.
Also Read:
- Agent Foundation Models: A Unified Approach to AI Problem Solving
- OpenAI Unveils gpt-oss-120b and gpt-oss-20b: New Open-Weight Models for Agentic AI
In conclusion, Datarus-R1 represents a significant step forward in training reasoning models. By focusing on complete problem-solving workflows and incorporating advanced training techniques, it delivers high accuracy and remarkable token efficiency, setting a new standard for AI in automated data analysis and broader problem-solving domains.
