--- title: KeenForgeAI emoji: 🔥 colorFrom: indigo colorTo: purple sdk: static pinned: false --- # 🔥 KeenForgeAI **Forging high-quality, niche vision datasets — and the open-source tool to build them.** KeenForgeAI is an open-source organization focused on **re-annotating and curating small, high-quality niche datasets** for computer vision, and on building **[KeenForge](https://github.com/KeenForgeAI/KeenForge)** — a local-first annotation & training desktop tool that makes this kind of careful work practical. 🔗 **GitHub:** · **Datasets:** --- ## 🛠 The tool: KeenForge > **CVPR 2026 Demo Track** · Formerly *AutoLabel Pro* KeenForge is a desktop application that lets you train your own object-detection model **without writing a single line of code**. Import images, draw a few boxes, and the AI trains itself in the background. **Your data never leaves your computer.** | | KeenForge | |---|---| | **Deployment** | Double-click an `.exe` (Windows) | | **Privacy** | **100% local & offline** — images are never uploaded | | **AI assistance** | Built-in **YOLO-World zero-shot** detection | | **Auto-training** | Triggers automatically after ~15 labels | | **Model ownership** | **You own your model** | | **Price** | **Free & open source (MIT)** | **Core features** - **Zero-shot cold start** — type any object name (`"welding defect"`, `"safety helmet"`, `"platelet"`) and KeenForge finds it immediately. No pre-training required. - **Self-looping active learning** — label ~15 images → auto-train in the background → the model suggests better boxes on the next images → repeat. The more you label, the smarter it gets. - **Infinite canvas** — smooth pan / zoom for large, high-resolution images. - **Export anywhere** — YOLO / COCO / VOC formats. 🎬 [Watch the demo](https://www.youtube.com/watch?v=shU62O8_40E) · ⭐ [Star on GitHub](https://github.com/KeenForgeAI/KeenForge) --- ## 🎯 Our mission: high-quality niche datasets Most public datasets are large but noisy. Smaller **well-annotated** datasets in *niche domains* are often the ones that actually move research forward — yet they are the hardest to find. We take public niche datasets — **industrial / manufacturing**, **medical & microscopy**, **agriculture**, **safety**, and more — and: 1. **Re-annotate them carefully** with a human-in-the-loop workflow in KeenForge (AI pre-labels → every image reviewed by a human → independent cross-check by a second annotator). 2. **Clean them** — tighten boxes, fix mislabels, add missed instances, remove duplicates. 3. **Document them** — full provenance, source datasets, correction reports and citations. 4. **Publish them openly** on **Hugging Face** and **ModelScope**. ### Published datasets | Dataset | Domain | Size | Description | |---|---|---|---| | [`KIT_BallScrew_Defects-corrected`](https://huggingface.co/datasets/KeenForgeAI/KIT_BallScrew_Defects-corrected) | 🔩 Industrial (ball screw) | 21,835 images · 0 boxes | Ball-screw surface defect image classification, 3 classes (ImageFolder) — cleaned, re-indexed and split version of KIT-BS-recls (CC BY-SA 4.0) | | [`TrafficSigns_YOLO-corrected`](https://huggingface.co/datasets/KeenForgeAI/TrafficSigns_YOLO-corrected) | 🚦 Transport (traffic signs) | 6,000 images · 6,000 boxes | Traffic sign detection, 6 classes (VOC + YOLO) — leak-safe re-split of the YOLOv8-YOLOv11 traffic sign benchmark dataset | | [`RoadCrack_DL_HY-corrected`](https://huggingface.co/datasets/KeenForgeAI/RoadCrack_DL_HY-corrected) | 🛣️ Civil (crack) | 2,189 images · 2,677 boxes | Pavement crack detection, 4 classes (VOC + YOLO) — cleaned box-level version of CrackDataset_DL_HY. Non-commercial research use only. | | [`VisuCeram_500x500-corrected`](https://huggingface.co/datasets/KeenForgeAI/VisuCeram_500x500-corrected) | 🧱 Industrial (ceramic) | 3,264 images · 6,042 boxes | Sanitary-ware ceramic surface defect detection, 7 classes (VOC + YOLO, Apache-2.0) — cleaned, de-folder and split version of VisuCeram 500x500 | | [`PV_Multi_Defect-corrected`](https://huggingface.co/datasets/KeenForgeAI/PV_Multi_Defect-corrected) | ☀️ Energy (PV) | 1,044 images · 3,403 boxes | Photovoltaic (PV) panel multi-defect detection, 5 classes (VOC + YOLO) — cleaned, de-duplicated and split version of the PV-Multi-Defect dataset | | [`DBA_Fire_Smoke-corrected`](https://huggingface.co/datasets/KeenForgeAI/DBA_Fire_Smoke-corrected) | 🔥 Safety (fire/smoke) | 3,846 images · 7,988 boxes | Fire and smoke detection, 2 classes (VOC + YOLO) — de-duplicated and split version of the DBA-Fire dataset | | [`CPLID_Insulators-corrected`](https://huggingface.co/datasets/KeenForgeAI/CPLID_Insulators-corrected) | ⚡ Energy (power-line) | 845 images · 1,566 boxes | Power-line insulator defect detection, 2 classes (VOC + YOLO) — cleaned and split version of the Chinese Power Line Insulator Dataset (CPLID) | | [`PlantDoc-corrected`](https://huggingface.co/datasets/KeenForgeAI/PlantDoc-corrected) | 🌱 Agriculture (plant disease) | 2,565 images · 8,887 boxes | Plant leaf disease detection, 29 classes (VOC + YOLO) — cleaned version of PlantDoc with train/test leakage removed and a reproducible split added | | [`PCB_AoI_KubeEdge-corrected`](https://huggingface.co/datasets/KeenForgeAI/PCB_AoI_KubeEdge-corrected) | 🖥️ Industrial (PCB) | 233 images · 1,011 boxes | PCB AOI defect detection, 2 classes (VOC + YOLO) — cleaned version of the KubeEdge-Ianvs PCB-AoI dataset, with a training-only augmentation set | | [`MagneticTile-corrected`](https://huggingface.co/datasets/KeenForgeAI/MagneticTile-corrected) | 🧲 Industrial (magnetic tile) | 1,318 images · 437 boxes | Magnetic tile surface defects, 5 classes (VOC + YOLO) — bounding-box standardized version of the Magnetic Tile Surface Defect Dataset | | [`PKU-Market-PCB-corrected`](https://huggingface.co/datasets/KeenForgeAI/PKU-Market-PCB-corrected) | 🔧 Industrial (PCB) | 693 images · 2,953 boxes | PCB defect detection, 6 classes (COCO + VOC + YOLO) — cleaned version of PKU-Market-PCB, with a board-disjoint split | | [`DeepPCB-corrected`](https://huggingface.co/datasets/KeenForgeAI/DeepPCB-corrected) | 🔌 Industrial (PCB) | 1,499 images · 10,004 boxes | PCB defect detection, 6 classes (TXT + VOC + YOLO) — cleaned version of DeepPCB (1,499 template/tested image pairs) | | [`NEU-DET-corrected`](https://huggingface.co/datasets/KeenForgeAI/NEU-DET-corrected) | 🔩 Industrial (steel) | 1,797 images · 4,177 boxes | Hot-rolled steel strip surface defects, 6 classes (VOC + YOLO) — cleaned version of the NEU Surface Defect Database | | [`GC10-DET-corrected`](https://huggingface.co/datasets/KeenForgeAI/GC10-DET-corrected) | 🏭 Industrial (steel) | 2,280 images · 3,542 boxes | Hot-rolled steel strip surface defects, 10 classes (Pascal VOC) — cleaned version of GC10-DET | | [`TXL-PBC-corrected`](https://huggingface.co/datasets/KeenForgeAI/TXL-PBC-corrected) | 🩸 Medical (hematology) | 1,256 images · 18,098 boxes | Peripheral blood cell detection (WBC / RBC / Platelets) — corrected version of TXL-PBC | | [`raccoon-corrected`](https://huggingface.co/datasets/KeenForgeAI/raccoon-corrected) | 🦝 Wildlife | 193 images · 211 boxes | Fully re-annotated version of the classic Raccoon detection dataset | *More coming — PCB, agriculture, safety, …* --- ## 💡 What we believe > **Only high-quality data can push visual AI forward.** Architectures are converging and compute keeps getting cheaper — but a model is only ever as good as the labels it learns from. We optimise for annotation quality, provenance and reproducibility, not raw size. --- ## 🤝 Join us KeenForgeAI is a community effort. You can help by: - **Use KeenForge** — report bugs or request features in the [issue tracker](https://github.com/KeenForgeAI/KeenForge/issues). - **Contribute code** — MIT-licensed; PRs, docs and translations welcome. - **Contribute datasets** — re-annotated a niche dataset? Open an issue and we'll help you publish it. - **Spread the word** — star the repo and share your workflow. The best vision datasets will come from many small communities working carefully, not from a few giant scrapes. **Come build with us.** --- ## 中文版 **KeenForgeAI 是一个开源组织，专注于「高质量小众视觉数据集」的重新标注与整理，并开发配套的本地化标注训练工具 [KeenForge](https://github.com/KeenForgeAI/KeenForge)。** ### 🛠 工具：KeenForge > **CVPR 2026 Demo Track** · 前身为 *AutoLabel Pro* KeenForge 是一款桌面应用，**无需写一行代码**就能训练你自己的目标检测模型：导入图片 → 画几个框 → AI 在后台自动训练。**数据全程不离开你的电脑。** - **零样本冷启动**：直接输入要检测的目标名称（如「焊接缺陷」「安全帽」「血小板」）， 内置 YOLO-World 立即识别，无需预训练 - **自循环主动学习**：标注约 15 张图 → 后台自动训练 → 模型在新图上给出更好的框 → 继续标注，越标越准 - **无限画布**：大图 / 高分辨率图片流畅缩放平移 - **一键导出**：YOLO / COCO / VOC 格式 - **100% 本地离线** · **模型归你所有** · **完全免费开源（MIT）** 🎬 [演示视频](https://www.youtube.com/watch?v=shU62O8_40E) ｜ ⭐ [GitHub 仓库](https://github.com/KeenForgeAI/KeenForge) ### 🎯 我们的使命 公开数据集往往「大而糙」——框松、漏标、错标、重复。真正推动研究的，常常是那些**小众领域里标注精良** 的小数据集，但它们恰恰最难找。 我们把公开的小众数据集（**工业制造**、**医疗显微**、**农业**、**安防**等）重新整理： 1. 用 KeenForge 的人机协同流程**重新精标**（AI 预标注 → 逐张人工复核 → 第二人交叉复核） 2. **清理**：收紧框、改错标、补漏标、去重复 3. **完整文档化**：来源、修正报告、引用信息齐全 4. **开源发布**到 **Hugging Face** 与 **ModelScope** **已发布数据集**：🔩 [`KIT_BallScrew_Defects-corrected`](https://huggingface.co/datasets/KeenForgeAI/KIT_BallScrew_Defects-corrected)（工业 — 滚珠丝杠表面质检，21,835 张 / 0 框）、🚦 [`TrafficSigns_YOLO-corrected`](https://huggingface.co/datasets/KeenForgeAI/TrafficSigns_YOLO-corrected)（交通 — 交通标志检测，6,000 张 / 6,000 框）、🛣️ [`RoadCrack_DL_HY-corrected`](https://huggingface.co/datasets/KeenForgeAI/RoadCrack_DL_HY-corrected)（土木 — 路面裂缝检测，2,189 张 / 2,677 框）、🧱 [`VisuCeram_500x500-corrected`](https://huggingface.co/datasets/KeenForgeAI/VisuCeram_500x500-corrected)（工业 — 陶瓷表面质检，3,264 张 / 6,042 框）、☀️ [`PV_Multi_Defect-corrected`](https://huggingface.co/datasets/KeenForgeAI/PV_Multi_Defect-corrected)（能源 — 光伏组件巡检，1,044 张 / 3,403 框）、🔥 [`DBA_Fire_Smoke-corrected`](https://huggingface.co/datasets/KeenForgeAI/DBA_Fire_Smoke-corrected)（安防 — 火焰与烟雾检测，3,846 张 / 7,988 框）、⚡ [`CPLID_Insulators-corrected`](https://huggingface.co/datasets/KeenForgeAI/CPLID_Insulators-corrected)（能源 — 电力巡线，845 张 / 1,566 框）、🌱 [`PlantDoc-corrected`](https://huggingface.co/datasets/KeenForgeAI/PlantDoc-corrected)（农业 — 植物病害，2,565 张 / 8,887 框）、🖥️ [`PCB_AoI_KubeEdge-corrected`](https://huggingface.co/datasets/KeenForgeAI/PCB_AoI_KubeEdge-corrected)（工业 — PCB，233 张 / 1,011 框）、🧲 [`MagneticTile-corrected`](https://huggingface.co/datasets/KeenForgeAI/MagneticTile-corrected)（工业 — 磁瓦表面，1,318 张 / 437 框）、🔧 [`PKU-Market-PCB-corrected`](https://huggingface.co/datasets/KeenForgeAI/PKU-Market-PCB-corrected)（工业 — PCB，693 张 / 2,953 框）、🔌 [`DeepPCB-corrected`](https://huggingface.co/datasets/KeenForgeAI/DeepPCB-corrected)（工业 — PCB，1,499 张 / 10,004 框）、🔩 [`NEU-DET-corrected`](https://huggingface.co/datasets/KeenForgeAI/NEU-DET-corrected)（工业 — 钢材表面，1,797 张 / 4,177 框）、🏭 [`GC10-DET-corrected`](https://huggingface.co/datasets/KeenForgeAI/GC10-DET-corrected)（工业 — 钢板表面，2,280 张 / 3,542 框）、🩸 [`TXL-PBC-corrected`](https://huggingface.co/datasets/KeenForgeAI/TXL-PBC-corrected)（医疗 — 血液学，1,256 张 / 18,098 框）、🦝 [`raccoon-corrected`](https://huggingface.co/datasets/KeenForgeAI/raccoon-corrected)（野生动物，193 张 / 211 框） ### 💡 我们的理念 > **只有高质量的数据集，才能推动视觉 AI 的发展。** 模型架构在趋同、算力越来越便宜，但模型的上限永远由标注质量决定。我们死磕的是标注质量、 可溯源性和可复现性，而不是数据量。 ### 🤝 加入我们 - **用 KeenForge**——提 issue / 提需求 - **贡献代码**（MIT 许可，欢迎 PR、文档、翻译） - **贡献数据集**——用 KeenForge 重新标注了小众数据集？开个 issue，我们帮你发布 - **帮忙传播**——给仓库点个 star，分享你的工作流 我们相信，未来最好的视觉数据集会来自许多小社区认真细致的工作，而不是少数几个巨大的爬取。 **欢迎一起共建 KeenForgeAI。** 🔗 GitHub: ｜ Hugging Face: