Fable-5.1-Max-Reasoning-Filtered-10000x is a high-quality multi-turn trajectory dataset for agentic coding and reasoning. It aims to use the code reasoning processes generated by the Fable 5.1 model under maximum reasoning intensity for model distillation and capability improvement.

The dataset contains 10,000 multi-turn trajectories, with a total of approximately 500,000,000 tokens and an average of about 50,000 tokens per trajectory. The data covers scenarios such as algorithm design, code implementation and performance optimization, complex debugging, multi-step logical reasoning, and adherence to complex prompt constraints.

Dataset Composition

• Algorithm design and code implementation: covers algorithm conception, program writing, and performance optimization, reflecting the complete process from task understanding to code implementation;
• Complex debugging and error handling: covers error localization, issue investigation, and exception handling, recording the multi-step analysis and correction process used to solve programming problems;
• Multi-step logical reasoning: covers complex problem decomposition, task planning, and staged solution, emphasizing the handling of multiple conditions and complex prompt constraints;
• Agentic interaction trajectories: include multi-turn messages and tool-related information, used to present the model’s interaction and problem-solving process in complex coding tasks.