laion/CoderForge-Preview-v6-1000

laion/CoderForge-Preview-v6-1000是togethercomputer/CoderForge-Preview数据集的一个子集，包含trajectories分割和filtered_reward1部分的数据，并转换为Qwen3兼容的think-first OpenHands-XML格式。v6版本在每个助手回合的开始注入了<think>REASONING</think>块，以解决模型训练中的问题。数据集中的每条记录包含轨迹ID、奖励、来源和消息列表，消息列表中的每个消息都有角色、内容和训练标志。数据集大小为1,000行，来源于1,386行，采用确定性随机抽样方法。

laion/CoderForge-Preview-v6-1000 is a row-subset of the togethercomputer/CoderForge-Preview dataset (from the trajectories split, filtered_reward1), rendered into Qwen3-compatible think-first OpenHands-XML wire format. The v6 version injects a <think>REASONING</think> block at the start of every assistant turn to address issues in previous versions. Each row in the dataset includes a JSON object with trajectory_id, reward, source, and a list of messages. Each message in the list has a role, content, and a train flag. The dataset size is 1,000 rows, sourced from 1,386 rows, using deterministic random sampling with seed=42.