Claude Fable 5 Agent Traces is a dataset of agent trajectories generated based on the Claude Fable 5 model. It is designed to provide validated, real-world agent trajectory data for instruction following, tool invocation, and supervised fine-tuning (SFT) of coded agents. This dataset contains 2,380 real-world conversation trajectories and 12,490 training samples. Each trajectory preserves the agent's complete causal exploration process, tool call parameters, execution results, error correction behavior, and final response. The sample tasks cover instruction compliance, tool invocation, code building, and debugging. The data has undergone runtime normalization, removing runtime pipeline code, UI decorations, control sequences, and verbose success output templates, while preserving the causal context.

Dataset composition:

• Tool calling: Contains 694 trajectories and 1,774 rows of training data.
• Instructions are as follows: Includes 559 trajectories and 1,436 rows of training data.
• Building: Contains 312 trajectories and 1,699 rows of training data.
• Debugging: Contains 280 trajectories and 1,424 rows of training data.
• Project & Integration: Contains 199 trajectories and 1,674 rows of training data.
• Clarification: Contains 101 trajectories and 187 rows of training data.
• Feature development: Includes 81 trajectories and 600 rows of training data.
• Error recovery: Contains 75 trajectories and 250 rows of training data.
• Seed authoring: Contains 49 trajectories and 3,272 rows of training data.
• Refactoring & Performance: Includes 21 trajectories and 112 rows of training data.
• VMware Cloud Foundation 9.1: Includes 5 tracks and 37 rows of training data
• Security: Includes 3 trajectories and 12 lines of training data.
• Uncategorized: Contains 1 trajectory and 12 rows of training data.

Data fields:

• task: Unique identifier for the task
• lang: data language or primary programming language
• category: Detailed task category
• Split: Dataset classification, comprising a training set (train) and a validation set (val).
• assistant_step: The round in which the agent responds, starting from 1.
• assistant_steps: The total number of rounds of agent responses in the source trajectory.
• target_message_index: The index of the final agent's message.
• n_messages: The total number of messages up to the target message.
• messages: The cumulative dialogue context up to the target message

Citation

If you use this dataset in your research, please cite the source: Claude Fable 5 Agent Traces — https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces