Skip to content

Dataset · greghavens

Kimi K3 traces for coding, tool use, and instructions

greghavens/kimi-k3-coding-and-debugging-traces

582 trajectories and 3,956 rows for distilling agentic coding and tool use.

What's inside

This dataset contains Kimi K3 behavioral trajectories covering instruction following, tool use, and agentic programming. It comprises 582 trajectories and 3,956 training rows, available in parquet (3 MB) and JSONL (72 MB). Generated with the open-source moonshiner harness to distill verified traces. Language: English.

What you can do with it

  • SFT for agentic coding models
  • LoRA fine-tuning for tool use
  • Evaluating instruction following and agentic trajectories
Tasks
Text generation
Languages
English
Size
1K–10K rows
Downloads this month
2.5K
Likes
68

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("greghavens/kimi-k3-coding-and-debugging-traces")