Papers with Code开源重建:把论文、代码和评测连起来
Papers with Code: An Open Research Catalog You Can Run
Papers with Code 是一个把论文、实现代码、模型、数据集与评测结果关联起来的研究目录。它解决的具体问题是:读完一篇论文后,怎样继续找到对应实现、实验对象与比较条件,而不用在多个网站之间从头寻找。
这里介绍 Hugging Face 组织下的 huggingface/paperswithcode,代码采用 Apache-2.0 许可。官方 README 将它描述为对旧 paperswithcode.com 的重建,线上入口是 paperswithcode.co;这份源码、当前站点和过去的历史服务需要分别理解,不能因为名字相同就假定数据完整继承、运行方式也完全相同。
Mycelium Protocol 核对的代码版本为 dc00d5ec4cca3f9810bfcf970126dc059b4a14e8。本文做了源码检查和八组 URL 归一化验证,没有运行完整站点、导入全库或测试搜索吞吐。
代码:https://github.com/huggingface/paperswithcode
它比一个论文收藏夹多了什么?
收藏夹保存的是地址,研究目录需要保存地址之间的关系。一篇论文可以对应多个代码实现;一个实现可能提供多个模型;同一个任务又可能在不同数据集、指标和数据划分上出现结果。只保留“论文标题加 GitHub 链接”,后续很难判断两个成绩能不能直接比较。
项目把论文、任务、方法、数据集、组织与作者等对象放在同一个目录中,还提供排行榜、论文阅读入口和公共 API。自动化任务负责导入研究内容、刷新关联信息与提出评测记录;这些功能的价值在于减少重复整理,让人把时间留给核查。
但“有关联”仍然只是起点。论文页出现一个仓库,不代表该仓库的某个提交已经复现了论文中的特定分数。使用排行榜时,仍应核对数据集版本、评测划分、指标方向、模型版本和原始实验来源。

检索与研究助手分别依赖什么?
README 列出的技术栈是 React 前端、FastAPI 后端和 PostgreSQL,数据库需要 vector 与 pg_trgm 扩展。源码还保留词法检索、向量检索以及混合排序路径,不能把“项目支持语义检索”理解为新装数据库后自然就拥有可用的语义索引。
以 PaperRetrievalService 为例,混合检索会合并词法与语义候选,并通过 RRF(倒数排名融合)形成结果。工程上至少要分清三件事:论文数据有没有导入、对应嵌入代次与向量索引有没有准备、查询端的嵌入服务有没有配置。网页能打开,只能说明其中一部分链路正常。
可选研究聊天通过目录工具取得材料,属于建立在目录之上的能力。基础目录与词法检索不应该和外部大模型调用混为一谈;启用聊天、语义嵌入或自动整理时,则需要按各自配置检查凭证、服务端点与成本。
这也是我们对自建顺序的建议:先让少量确定的论文可导入、可查回、关联正确,再补向量检索和助手。这样出现错误时,能区分是缺数据、关联错误,还是检索与模型层的问题。
本地实际验证了什么?
我们直接加载该版本的 backend/utils/github_utils.py,调用上游的 normalize_github_url,运行八组输入检查:
| 输入情况 | 本地结果 |
|---|---|
| 所有者与仓库名大小写不同 | 归一为小写仓库身份 |
| 省略协议的 GitHub 地址 | 补全为标准 HTTPS 地址 |
克隆地址带 .git | 去掉克隆后缀 |
链接指向 tree/main/backend | 仍归到仓库本身 |
| 链接附带查询参数 | 不产生新的仓库身份 |
| 非 GitHub 域名 | 返回无效结果 |
| 空字符串 | 返回无效结果 |
| 只有组织名、没有仓库名 | 返回无效结果 |
八组检查全部通过。它们验证的是一个具体组件,不是完整项目的质量认证,更没有测量检索准确率或部署性能。
这个组件虽小,却对应目录系统的真实问题:同一个实现经常以 README 链接、克隆地址和目录链接出现。先把它们归到相同身份,后续才有可靠的去重与关联基础。即便如此,不同分支、提交、模型权重和评测配置仍需要单独记录;归一仓库不能替代固定实验版本。
自建时,最容易低估哪一部分?
官方开发说明要求 Python 3.12、Node.js 18 以上及 PostgreSQL 16,并提供 uv 安装、数据库初始化和前后端启动步骤。初始化后的数据库是空的,还需要使用维护 CLI 导入论文。
一个小规模试用可以从官方给出的维护命令开始:
uv run --env-file backend/keys.env pwc paper index 1706.03762
uv run --env-file backend/keys.env pwc paper info 1706.03762
uv run --env-file backend/keys.env pwc search "vision-language models"
这些命令以已安装依赖并完成数据库配置为前提,不是复制三行就能得到完整站点。Hugging Face 与 GitHub 凭证可用于更高的导入限额;研究助手及托管集成还有各自的配置。项目也包含迁移、调度和部署资料,自建者需要为这些长期维护工作留出精力。
代码的 Apache-2.0 许可同样不等于它链接到的所有论文、数据集和模型都采用相同许可。研究目录帮助找到材料,材料本身的授权与使用条件仍应沿原始来源核对。
谁值得试用,怎样判断是否有用?
经常整理细分领域论文的研究者、需要共享技术证据的团队,以及想搭建内部研究目录的工程师,都可以先用一小组熟悉的论文试用。验收重点可以很具体:能否查到正确实现,能否区分不同版本,评测记录是否保留来源,以及错误关联能否被发现并修正。
对只想每天浏览新论文的人,先使用公开站点通常更省事。对要维护自己领域证据链的人,这份仓库更值得借鉴的,是把导入、身份归一、关联、检索和维护拆成可检查环节的做法。目录可以缩短找到证据的路,实验是否成立,最终仍要回到可执行的实现和明确的评测条件。
相关工程阅读:https://blog.mushroom.cv/blog/awesome-3d-coding-research-map-evaluation-guide/
源码与许可:https://github.com/huggingface/paperswithcode/tree/dc00d5ec4cca3f9810bfcf970126dc059b4a14e8
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用,须注明作者姓名及原文链接,不得去除署名后以原创发布。
Papers with Code connects research papers with implementations, models, datasets and evaluation records. Its practical purpose is to make the next step after reading a paper easier: locating the implementation and the conditions behind a reported result.
This article covers huggingface/paperswithcode, licensed under Apache-2.0, at commit dc00d5ec4cca3f9810bfcf970126dc059b4a14e8. The README describes a reconstruction of the former paperswithcode.com and points to paperswithcode.co as the live site. The current codebase, current hosted service and historical service should be distinguished rather than assumed to have identical data or behavior.
Mycelium Protocol inspected the source and ran eight checks against an upstream URL utility. We did not deploy the complete application, import its full catalog or benchmark search performance.
Repository: https://github.com/huggingface/paperswithcode
What does a research catalog add to bookmarks?
A bookmark stores an address. A useful research catalog also stores relationships: a paper may have several implementations; an implementation may publish several models; and results for the same task may use different datasets, splits and metrics.
This project organizes papers alongside tasks, methods, datasets, organizations and authors. It also provides leaderboards, a paper reader and a public API. Scheduled jobs import research and refresh links or propose evaluation records. These tools reduce repeated collection work, leaving more time for checking the evidence.
A linked repository is still not proof that a particular commit reproduces a particular score. Readers should verify the dataset version, evaluation split, metric direction, model version and original experiment source before comparing results.

What do search and the optional assistant require?
The documented stack combines a React frontend, a FastAPI backend and PostgreSQL with vector and pg_trgm extensions. The source includes lexical, semantic and hybrid retrieval paths. Supporting semantic search does not mean that a newly initialized database already contains usable embeddings.
For example, PaperRetrievalService merges lexical and semantic candidates using reciprocal rank fusion. Operators must distinguish imported paper data, a prepared embedding generation and index, and a configured query embedding service. A working web page does not establish that all these components are ready.
Optional research chat uses tools grounded in the catalog. The basic catalog and lexical search should be distinguished from external model calls. Chat, embeddings and automated curation have their own configuration, credentials and operating costs.
Our suggested deployment order is to import a small set of known papers, confirm retrieval and associations, and then add embeddings and the assistant. This makes missing data easier to distinguish from a retrieval or model problem.
What did we actually test?
We loaded the upstream backend/utils/github_utils.py directly and called normalize_github_url with eight inputs. The checks covered mixed-case repository names, missing URL schemes, .git suffixes, directory links and query parameters, plus rejection of another domain, an empty input and an organization-only address. All eight checks passed.
This is a component check, not an assessment of the whole application. It establishes neither search accuracy nor deployment performance. Its relevance is concrete: the same implementation often arrives as a README link, clone URL or directory link. Canonical repository identity gives deduplication and association a more dependable starting point.
Branches, commits, model weights and evaluation configurations still need separate tracking. Normalizing a repository address does not pin an experiment.
What should self-hosters budget for?
The development instructions specify Python 3.12, Node.js 18 or later and PostgreSQL 16. They cover dependency installation with uv, database initialization and frontend/backend startup. The initialized schema starts empty; the maintenance CLI is used to index papers.
After configuring dependencies and the database, the documented commands can index a known paper, read it back and perform a search:
uv run --env-file backend/keys.env pwc paper index 1706.03762
uv run --env-file backend/keys.env pwc paper info 1706.03762
uv run --env-file backend/keys.env pwc search "vision-language models"
These commands are not a complete deployment recipe on their own. Hugging Face and GitHub credentials can increase import rate limits; chat and hosted integrations need additional settings. Migrations, scheduled jobs and deployment operations also require ongoing attention.
Apache-2.0 applies to this codebase, not automatically to every linked paper, dataset or model. Follow original sources to check the conditions for the artifacts you intend to use.
Who should try it?
Researchers maintaining a focused literature collection, teams sharing technical evidence and engineers building an internal catalog can start with a small set of familiar papers. Evaluate whether implementations are correctly linked, versions remain distinguishable, evaluation sources are preserved and incorrect associations can be corrected.
For daily browsing, the hosted service may be sufficient. For building a domain-specific evidence collection, the more useful lesson is the separation of import, identity normalization, associations, retrieval and maintenance into inspectable steps. A catalog shortens the route to evidence; reproducibility still depends on executable implementations and explicit evaluation conditions.
Related reading: https://blog.mushroom.cv/blog/awesome-3d-coding-research-map-evaluation-guide/
Pinned source: https://github.com/huggingface/paperswithcode/tree/dc00d5ec4cca3f9810bfcf970126dc059b4a14e8
Retrieval implementation: https://github.com/huggingface/paperswithcode/blob/dc00d5ec4cca3f9810bfcf970126dc059b4a14e8/backend/services/paper_retrieval_service.py
URL component: https://github.com/huggingface/paperswithcode/blob/dc00d5ec4cca3f9810bfcf970126dc059b4a14e8/backend/utils/github_utils.py
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论