Simonkami/BvsALL_explained_selftrain

BvsALL_explained_selftrain_round1_half_original是一个本地混合的监督微调（SFT）数据集，专为自然语言处理任务设计。它由两个来源组合而成：一是原始数据集Simonkami/BvsALL_explained的train分割部分，二是自训练文件selftrain_good_samples_round1.jsonl。在混合过程中，原始数据使用了1699条样本（占原始总数8499条的20%），自训练数据使用了6796条样本（从原始13490条中经过有效性验证和动作平衡后选择9674条，再进一步筛选）。目标自训练比例为0.8，实际达到0.8000，最终数据集总共有8495条样本。数据列包括instruction（指令）、input（输入）和output（输出），用于训练模型遵循指令并生成响应。此外，还提供了一个包含额外元数据（如source和score字段）的版本train_with_meta.jsonl供检查，但推荐使用train.jsonl进行实际的SFT训练。BvsALL_explained_selftrain_round1_half_original is a local mixed supervised fine-tuning (SFT) dataset built from two sources: the original dataset Simonkami/BvsALL_explained split train and a self-training file selftrain_good_samples_round1.jsonl. The mixing settings involve using 1699 samples from the original dataset (20% of the original total of 8499) and 6796 samples from the self-training data (selected from 13490 raw samples after validation and action balancing to 9674, then further filtered). The target self-training ratio is 0.8, with an actual ratio of 0.8000, resulting in a final total of 8495 samples. The dataset columns include instruction, input, and output, designed for training models to follow instructions and generate responses. An additional file train_with_meta.jsonl contains extra fields such as source and score for inspection, but train.jsonl is recommended for SFT training.