SheatNoisette/vlang-poc-dataset
收藏资源简介:
--- license: mit language: - en tags: - code pretty_name: Proof-Of-Concept Vlang dataset size_categories: - n<1K --- # Vlang Code Dataset This dataset was a proof-of-concept for code finetuning on [vlang.io](vlang) code as a Q&A format. It contains partial examples code from: - V Compiler examples - V Coreutils - vglyph - vsdl - vsl - vgui Compatible with Unsloth Studio Beta. Most of the Q&A were written by GPT-OSS-120b. A second pass was done on the "vdoc.md" vdocumentation with Minimax's Minimax-2.7. ## Content The dataset contains three columns: - type: Where the data comes from - question: The synthetic question - answer: A LLM friendly anwser The possible values for types are: - manual: Questions written by hand and rewritten (V compiler, CLI, trivia, v help) - vdoc: The vdoc.md present in the V compiler repository, second pass with Minimax-2.7 - vlang_examples: Some examples from the V compiler - vsl_examples: Examples from the VSL - vgui_examples: Example from VGui - vglyph_examples: Examples from vglyph - vcoreutils: Some one file V Coreutils - vsdl_examples: Some vsdl examples ## Example Details Here is a list of partial examples import. V Compiler examples (git 58f18573557247e9c743f7d60e5b3ab783f74479): - v/examples/*.v - v/examples/gg - v/examples/database - v/examples/linear_regression/ - v/examples/word_counter/ - v/examples/smtp - v/examples/thread_safety/ - v/examples/web_crawler - v/examples/graphs - v/examples/sokoban/ - v/examples/get_weather/ - v/examples/concurrency/ - v/examples/hot_reload/ - v/examples/process/ - v/examples/sokol/01_cubes/ - v/examples/term.ui/ - v/examples/orm/ - v/examples/compiletime VSDL (git c7b4a451e4aec2dd7538258065c8ad1c8b9df466): - vsdl/examples/versions/main.v - vsdl/examples/version/main.v - vsdl/examples/tvintris/tvintris.v - vsdl/examples/basic_window/main.v - vsdl/examples/basic_mixer/basic_mixer.v - vsdl/examples/basic_image/basic_image.v VCoreutils (git ac8118538b0a5f7cec68d668923f3c21479f85a6): - arch.v - b2sum.v - base64.v - basename.v - cat.v - comm.v - cut.v - dirname.v - echo.v - expand.v - expr.v - factor.v - false.v - fmt.v - fold.v - groups.v - head.v - id.v - left_bracket.v - link.v - logname.v - ls.v - md5sum.v - mkdir.v - nl.v - printenv.v - printf.v - pwd.v - seq.c.v - sha1sum.v - sha224sum.v - sha256sum.v - sha384sum.v - sha512sum.v - shuf.v - sleep.v - tac.v - test.c.v - true.v - uname.v - uptime.c.v - wc.v - yes.v vsl (git 11b57d4e77ffd4bbd4893ee99e3d74ff372ebac8): - examples/*.v ## Known limitations There’s a list of problems I observed when testing fine-tuned models on my dataset: - Underrepresentation of flagparser usage and many other modules, leading to severe hallucinations on vlib - Missing examples of V compiler corner cases and syntax, models may get small bits of syntax wrong - Asking about CLI apps may cause the import of the "common" modules from Coreutils, which don't exist on vlib - Poor data on V compiler CLI usage - Short and rigid explanations - Unfiltered code: models may hallucinate authors in code headers, add useless comments, and insert "TODOs" - No examples of multi-file code and modules, leading to an inability to create multi-file programs - Poor understanding of tests - Bias toward writing long files - Limited knowledge of the libraries used: models may generate code for a library without being able to explain what the library does/is - No CoT available - Q&A format - Some responses clarification contains hallucination ## LLMs used GPT-OSS-120b was used for a cost saving measure at the risk of high hallucination. Minimax-2.7 for a second pass as one of "best" available Hallucination/Capability/Cost best overall on OpenRouter (Source: artificial-analysis.ai, AA-Omni test) ## Cost LLMs used were from OpenRouter, no training policy. These numbers includes testing. GPT-OSS-120b (non exacto) providers: - Clarifai - Google Vertex - DeepInfra - SiliconFlow Minimax 2.7 providers: - AtlasCloud - MiniMax Other data: - Token used (GPT-OSS-120b 1.95M + Minimax 2.7 355K): 2.31M - Cost: 0.658$ - API Calls number: GPT-OSS 986, Minimax 160 ## License The projects scrapped at the time were under the MIT License. I do not claim ownership on any of the scrapped code.
许可证:MIT 语言: - 英语 标签: - 代码 展示名称:V语言概念验证数据集 大小类别: - n<1K # V语言代码数据集 本数据集旨在作为基于[vlang.io](vlang)代码的问答格式代码微调的概念验证项目。 其包含源自以下来源的部分示例代码: - V编译器示例 - V核心工具集(V Coreutils) - vglyph - vsdl - vsl - vgui 本数据集兼容Unsloth Studio Beta版本。 绝大多数问答内容由GPT-OSS-120b撰写,随后针对V编译器仓库中的`vdoc.md`文档,使用Minimax的Minimax-2.7进行了二次优化与校验。 ## 数据内容 本数据集包含三列数据: - `type`:数据来源类别 - `question`:合成式问题 - `answer`:适配大语言模型(Large Language Model, LLM)的回答 `type`字段的可选取值如下: - `manual`:手写并经改写的问题(涵盖V编译器、命令行界面(Command Line Interface, CLI)、常识问答、v帮助文档相关内容) - `vdoc`:源自V编译器仓库的`vdoc.md`文档,经Minimax-2.7二次处理 - `vlang_examples`:V编译器自带的部分示例代码 - `vsl_examples`:VSL的示例代码 - `vgui_examples`:VGui的示例代码 - `vglyph_examples`:vglyph的示例代码 - `vcoreutils`:单文件形式的V核心工具集代码 - `vsdl_examples`:vsdl的部分示例代码 ## 示例详情 以下为部分示例代码的导入清单: V编译器示例(对应Git提交哈希`58f18573557247e9c743f7d60e5b3ab783f74479`): - `v/examples/*.v` - `v/examples/gg` - `v/examples/database` - `v/examples/linear_regression/` - `v/examples/word_counter/` - `v/examples/smtp` - `v/examples/thread_safety/` - `v/examples/web_crawler` - `v/examples/graphs` - `v/examples/sokoban/` - `v/examples/get_weather/` - `v/examples/concurrency/` - `v/examples/hot_reload/` - `v/examples/process/` - `v/examples/sokol/01_cubes/` - `v/examples/term.ui/` - `v/examples/orm/` - `v/examples/compiletime` VSDL(对应Git提交哈希`c7b4a451e4aec2dd7538258065c8ad1c8b9df466`): - `vsdl/examples/versions/main.v` - `vsdl/examples/version/main.v` - `vsdl/examples/tvintris/tvintris.v` - `vsdl/examples/basic_window/main.v` - `vsdl/examples/basic_mixer/basic_mixer.v` - `vsdl/examples/basic_image/basic_image.v` VCoreutils(对应Git提交哈希`ac8118538b0a5f7cec68d668923f3c21479f85a6`): - `arch.v` - `b2sum.v` - `base64.v` - `basename.v` - `cat.v` - `comm.v` - `cut.v` - `dirname.v` - `echo.v` - `expand.v` - `expr.v` - `factor.v` - `false.v` - `fmt.v` - `fold.v` - `groups.v` - `head.v` - `id.v` - `left_bracket.v` - `link.v` - `logname.v` - `ls.v` - `md5sum.v` - `mkdir.v` - `nl.v` - `printenv.v` - `printf.v` - `pwd.v` - `seq.c.v` - `sha1sum.v` - `sha224sum.v` - `sha256sum.v` - `sha384sum.v` - `sha512sum.v` - `shuf.v` - `sleep.v` - `tac.v` - `test.c.v` - `true.v` - `uname.v` - `uptime.c.v` - `wc.v` - `yes.v` vsl(对应Git提交哈希`11b57d4e77ffd4bbd4893ee99e3d74ff372ebac8`): - `examples/*.v` ## 已知局限性 在使用本数据集微调模型的测试过程中,笔者发现了以下若干问题: 1. 对flagparser等众多模块的覆盖不足,导致模型在涉及vlib库时极易产生幻觉内容 2. 缺少V编译器边缘场景与语法的示例,模型可能在部分语法细节上出现错误 3. 当询问CLI应用相关问题时,模型可能会错误导入核心工具集中不存在的`common`模块 4. V编译器CLI使用场景的相关数据不足 5. 解释内容简短且格式僵硬 6. 代码未经过滤处理:模型可能会在代码头部虚构作者信息、添加无用注释或插入`TODO`标记 7. 无多文件代码与模块相关示例,导致模型无法生成多文件程序 8. 对测试相关内容的理解能力有限 9. 存在生成过长文件的生成偏向 10. 对所用库的认知不足:模型可能生成某库的代码,但无法解释该库的功能或用途 11. 未提供思维链(Chain of Thought, CoT)相关示例 12. 数据集整体采用问答格式 13. 部分回复的澄清内容存在幻觉现象 ## 所用大语言模型 为控制成本,初期优先使用GPT-OSS-120b,但该模型存在较高的幻觉风险。随后使用Minimax-2.7进行二次校验,该模型是OpenRouter平台上综合表现(幻觉控制、模型能力、成本)较佳的模型之一(数据来源:artificial-analysis.ai的AA-Omni测试结果)。 ## 成本统计 所用大语言模型均来自OpenRouter平台,未涉及特定训练策略,以下统计数据包含测试环节的开销。 GPT-OSS-120b(非精确统计)的服务提供商包括: - Clarifai - Google Vertex - DeepInfra - SiliconFlow Minimax 2.7的服务提供商包括: - AtlasCloud - MiniMax 其他统计数据: - 总Token消耗量(GPT-OSS-120b 1.95M + Minimax 2.7 355K):2.31M - 总开销:0.658美元 - API调用次数:GPT-OSS 986次,Minimax 160次 ## 许可证 本次爬取的项目当时均采用MIT许可证,本人不对所爬取的任何代码主张所有权。





