sob-ft-json-training
收藏资源简介:
SOB-FT-JSON-Training 是一个从 ChristianAzinn/json-training 映射而来的合成、无来源文档的 JSON 提取指令微调数据集。原始数据集中包含 query、schema 和 response 三个字段,本数据集将其转换为 SOB 微调格式(移除所有空列),具体对应关系为:question 取自 query,json_schema 取自 schema,ground_truth 取自 response。由于原始数据为合成生成且无来源文档,context 字段被设为 null 并已删除。数据规模方面,原始 20,644 条经过 JSON 语法验证和字符长度不超过 12,000 的过滤后,最终保留 19,981 条有效样本。每条样本还包含 question_type(固定为 synthetic_ungrounded)和 source_dataset(标识来源)两个标签。该数据集主要用于增强模型根据给定 JSON Schema 生成正确 JSON 输出的能力,而非提升基于文档的精确提取性能。训练时建议对该数据集采用较低的采样权重。
SOB-FT-JSON-Training is a synthetic, source-free document JSON extraction instruction fine-tuning dataset mapped from ChristianAzinn/json-training. The original dataset contains three fields: query, schema, and response. This dataset converts them into the SOB fine-tuning format (removing all empty columns), with the specific mapping: question is taken from query, json_schema from schema, and ground_truth from response. Since the original data is synthetically generated without source documents, the context field is set to null and removed. In terms of data scale, after filtering the original 20,644 entries through JSON syntax validation and character length not exceeding 12,000, 19,981 valid samples are retained. Each sample also includes two labels: question_type (fixed as synthetic_ungrounded) and source_dataset (indicating source). This dataset is mainly used to enhance the models ability to generate correct JSON output based on a given JSON Schema, rather than improving document-based precise extraction performance. It is recommended to use a lower sampling weight for this dataset during training.
SOB-FT-JSON-Training 数据集概述
基本信息
- 许可证: MIT
- 语言: 英语 (en)
- 标签: json-extraction、synthetic、sob、fine-tuning
数据来源
该数据集是基于 ChristianAzinn/json-training 映射生成的独立语料库,属于合成数据,主要用于 JSON 提取任务的微调。
与源数据的差异
| 维度 | 源数据集 | 本数据集 |
|---|---|---|
| 格式 | query、schema、response |
SOB FT 格式(已移除空列) |
| 基础依据 | 合成 / 无依据 | context=null(无源文档) |
| 问题字段 | query |
question=query |
| 模式字段 | schema |
json_schema(其中 481 行存在无效源架构) |
| 目标字段 | response |
ground_truth |
| 行数 | 20,644 条源数据 → 19,981 条映射后数据 | 经 JSON 验证及字符数 ≤12k 过滤 |
| 标签 | 无 | question_type=synthetic_ungrounded |
| 来源标注 | 无 | source_dataset=ChristianAzinn/json-training |
数据特点
- 所有全空列(包括
context)已被删除 - 训练建议:应保持较低的采样权重——该数据适用于架构增强,而非基于文档的抽取准确性提升
复现与合并命令
- 复现:
python3 merge_json_training.py --dry-run和python3 merge_json_training.py - 合并到提取流程:
python3 merge_ft_corpus.py --corpus mariem123kfg/sob-ft-json-training --source-prefix ChristianAzinn/json-training




