PGCodeLLM/

ccr-bench

	v2_human_gpt — SFT Training Data for Qwen3-Coder




	
	Overview

This directory contains SFT training data converted to a format that works with stock LLaMA-Factory (no modifications) and produces identical tokens to what the model sees at inference time through the Qwen3-Coder Jinja2 chat template.

per_turn_plan_qwen3coder.jsonl — 25,033 per-turn plan samples

	Why This Format




	
	The Problem

Qwen3-Coder uses a native XML tool call format that… See the full description on the dataset page: https://huggingface.co/datasets/PGCodeLLM/ccr-bench.