gene-general-ml-instruct
收藏资源简介:
general-ml-instruct v10是一个用于通用机器学习的门控通过指令数据集,包含230条高质量合成记录。该数据集专为文本生成和问答任务设计,采用CC-BY-4.0许可证发布。数据通过Gene管道系统生成,该系统采用来源追溯优先的方法:从ArXiv、GitHub和Hugging Face(仅限宽松许可证)抓取来源材料,合成示例需通过严格的六阶段质量门控流程,包括生成、批判与修订编辑、LLM评判、对抗性二次评判、证据验证(每条保留的数据对都包含可证明出现在其来源中的引用)以及代码片段的沙箱执行。数据集包含14种不同的任务类型,主要包括概念问答(49条)、特征提取(24条)、使用问题(24条)、实现草图(24条)、代码解释(21条)、问题陈述(16条)、摘要(14条)、比较(12条)、方法解释(11条)、故障排除(10条)、标题生成(10条)、局限性分析(7条)、未来工作(5条)和通用任务(3条)。所有记录都经过LLM评判,平均得分0.975(满分1.0),且全部携带已验证的来源引用。数据集通过manifest.json文件确保完全可复现性,能够通过指定命令字节级重建。
general-ml-instruct v10 is a quality-gated instruction dataset for general machine learning (general-ml), containing 230 high-quality synthetic records. This dataset is specifically designed for text generation and question answering tasks, and is released under the CC-BY-4.0 license. The dataset is generated via the Gene pipeline system, which adopts a source-trace-first methodology: it scrapes source materials from ArXiv, GitHub, and Hugging Face (with permissive licenses only), and synthetic examples must pass a rigorous six-stage quality gated process, including generation, critique and revision editing, LLM judging, adversarial secondary judging, evidence verification (each retained data pair includes verifiable citations appearing in its source material), and sandbox execution of code snippets. The dataset covers 14 distinct task types, mainly including conceptual QA (49 entries), feature extraction (24 entries), usage queries (24 entries), implementation sketches (24 entries), code explanation (21 entries), problem statement (16 entries), summarization (14 entries), comparison (12 entries), method explanation (11 entries), troubleshooting (10 entries), title generation (10 entries), limitation analysis (7 entries), future work (5 entries), and general tasks (3 entries). All records have been evaluated by LLMs, with an average score of 0.975 (out of 1.0), and all carry verified source citations. The dataset ensures full reproducibility via the manifest.json file, and can be reconstructed at the byte level through specified commands.
数据集概述
- 数据集名称:general-ml-instruct v11
- 许可证:CC-BY-4.0
- 语言:英语(en)
- 任务类型:文本生成、问答
- 数据集规模:n<1K(共280条记录)
- 创建时间:2026-06-12T19:20:20+00:00
- SHA-256指纹:
ab659aaf5d71f5db0e450bb72aa9dac8cf08da6d59b52f80216321132ab9edfa
数据来源与生成
- 种类:合成数据(synthetic)
- 领域:通用机器学习(general-ml)
- 生成模型:Qwen3-4B-Instruct-2507-Q4_K_M.gguf(后端:llama)
- 生成管线:Gene 管线(v2.0.0),一个以来源追溯为先的训练数据管线,数据来源包括 ArXiv、GitHub 和 Hugging Face(仅使用宽松许可证来源)。
- 生成过滤条件:
{"min_quality": 0.55, "limit": 1000, "source": null, "backend": "llama", "min_judge": 0.7}
质量门控与评估
- 质量门控流程:经过六阶段门控——生成、批评与修订的编辑环节、LLM 评判、对抗性二次评判、证据验证(每个保留的配对都包含一条可证明来源于其源文本的引用)、以及代码的沙盒执行。
- 评判分数:280/280 条记录均经过评判,平均分 0.975,最低 0.784,最高 1.000(基于扎实性、有用性、清晰性的评分标准,并带有可验证的源引用)。
- 门控通过情况:
- 280/280 条记录携带可验证的源引用
- 167/280 条记录通过了对抗性评审
- 107 条记录经过编辑环节改进
- 0/6 条代码片段在沙盒中成功执行
任务类型分布
| 任务类型 | 数量 |
|---|---|
| conceptual-qa (概念问答) | 54 |
| implementation-sketch (实现草图) | 33 |
| feature-extraction (特征提取) | 31 |
| code-explanation (代码解释) | 30 |
| usage-question (使用问题) | 30 |
| troubleshooting (故障排除) | 17 |
| comparison (比较) | 17 |
| problem-statement (问题陈述) | 16 |
| summarization (摘要) | 14 |
| method-explanation (方法解释) | 11 |
| title-generation (标题生成) | 10 |
| limitations-analysis (局限性分析) | 7 |
| future-work (未来工作) | 5 |
| general (通用) | 5 |
数据来源与可复现性
- 数据文件
data.jsonl中的每条记录都携带其来源/出处信息。 manifest.json文件记录了确切的记录 ID,通过命令gene rebuild --manifest manifest.json可字节完全一致地重新生成该数据集(SHA-256 验证)。





