Kimi K3 Coding & Debugging traces (full name Kimi K3 Coding, Tool Use & Instruction Following Traces) is a dataset of instruction following, tool invocation, and programming traces extracted from a Kimi K3 model. It aims to perform supervised fine-tuning (SFT) on instruction execution, tool invocation, and the coding agent. This dataset contains 3,906 rows of training data from 577 complete trajectories. Tasks cover categories such as system construction and implementation, defect diagnosis and debugging, tool invocation and orchestration, project migration and integration, error recovery, security hardening, and feature development. All data has undergone runtime normalization, removing redundant information such as UI decorations and control sequences, retaining only causal context, and has been validated through deterministic testing and independent review.

Dataset composition

• System construction and implementation: approximately 308 trajectories, 2,058 rows of training data
• Defect diagnosis and debugging: Approximately 149 tracks, 949 lines of training data
• Tool usage and orchestration: Approximately 79 trajectories, 636 rows of training data
• Project migration and integration: Approximately 26 trajectories, 175 rows of training data
• Bug fixes: Approximately 5 trajectories, 31 rows of training data
• Security reinforcement: Approximately 2 trajectories, 12 lines of training data
• Reconstruction and optimization: Approximately 2 trajectories, 8 rows of training data
• Feature development: Approximately 2 trajectories, 15 lines of training data
• Other tasks: Approximately 4 trajectories, 23 lines of training data

Data fields:

• task: A stable unique identifier for a task
• lang: the primary language of data processing or programming language
• category: Detailed task category
• split: A marker indicating that trajectories do not intersect.
• assistant_step / assistant_steps: Current target step count and total source trajectory steps.
• target_message_index / n_messages: Target assistant message index and total number of messages.
• messages: A list of accumulated conversation context objects stored as JSON objects, ending with the target assistant's reply.

Citation

If you use a dataset in your research, please cite the data source: Kimi K3 Coding, Tool Use & Instruction Following Traces — https://huggingface.co/datasets/greghavens/kimi-k3-coding-and-debugging-traces