setground-kilogram-tasks
收藏资源简介:
SetGround KiloGram 元数据是一个面向 KiloGram 自然 whole/part 标注的元数据数据集,README 元信息显示其语言为英语,任务类别为 other,仓库许可为 MIT。该数据集只包含 KiloGram 自然 whole/part 标注的无身份字段投影、来源校验和划分盘点;不包含图像、参与者 ID、时间或操作轨迹,也不包含本项目训练模型或交互收益结果。上游来源为 lil-lab/kilogram 仓库(版本 f2cd80201eedc5f4882bd88ba32de39195d9e9f3)及论文 arXiv:2211.16492。上游 README 明确:除 Tangram 图像外采用 MIT;版权及许可全文随 SOURCE_README.md 保留。图像来自书籍扫描件,另受教育科研用途说明约束,因此不能将本仓库 MIT 标签套用于图像。数据快照为 snapshots/2026-10-24-source-inventory-v1。annotations.jsonl.gz 逐行包含 source(full/dense)、shape、annotation_index、whole、part(七个 polygon ID 到自然部件名)等字段。full 和 dense 两个来源分别保留,重复标注不合并,不应简单相加当作独立样本。splits.json 保存官方旧 ID 划分与官方 documented_aliases,不提供重新划分。summary.json 保存来源 blob/SHA256、完整计数、冲突、成本与源码绑定;config.json 固定五个源文件。SOURCE_README.md 和 DATASET_README.md 为原版许可/数据说明。当前 full 为 1013 个形状、10173 条标注;dense 为 74 个形状、3971 条标注。旧 split 有 1016 个 ID,按官方别名合并后为 1013 个;跨 split 组件为 0 个,具体列表见 summary。README 指出计数不证明视觉独立性、候选组内歧义或追问收益。复现方面,代码位于 Git 仓库,inventory 源码版本为 83c8091,发布源码版本为 36db885c1e785733f43592d4a6eef81711e253b3。可使用 Python 3.10 标准库运行 python code/audit_kilogram_inventory.py --config code/configs/kilogram_inventory.json --output <persistent-output> --code-revision 83c8091。上游 JSON 只在内存中读取,输出是字段白名单投影;完整原始文件不在快照中。固定来源与新输出的 SHA256 见 summary,所有发布文件按此次不可变 HF revision 回下载校验。本轮无图像下载、VLM 调用、学习策略评测或 RL 训练。拟研究多个目标共享候选且受询问成本约束的匹配;本数据本身仍是静态标注,而不是反事实回答环境。
SetGround KiloGram metadata is a metadata dataset for KiloGram natural whole/part annotations. According to the README metadata, its language is English, task category is other, and repository license is MIT. The dataset only contains identity-free field projections of KiloGram natural whole/part annotations, source verification, and split inventory; it does not include images, participant IDs, time or operation traces, nor trained models or interaction benefit results from this project. Its upstream source is the lil-lab/kilogram repository (version f2cd80201eedc5f4882bd88ba32de39195d9e9f3) and paper arXiv:2211.16492. The upstream README states that MIT applies except for Tangram images; the full copyright and license text is retained in SOURCE_README.md. Images come from book scans and are also constrained by an educational/scientific-use statement, so this repositorys MIT label must not be applied to images. The data snapshot is snapshots/2026-10-24-source-inventory-v1. annotations.jsonl.gz contains line-by-line fields including source (full/dense), shape, annotation_index, whole, part (seven polygon IDs to natural part names). The full and dense sources are retained separately; duplicate annotations are not merged and should not simply be added as independent samples. splits.json stores the official old ID splits and official documented_aliases, and does not provide re-splits. summary.json stores source blob/SHA256, complete counts, conflicts, cost, and source-code binding; config.json fixes five source files. SOURCE_README.md and DATASET_README.md are the original license/data documentation. Currently full has 1013 shapes and 10173 annotations; dense has 74 shapes and 3971 annotations. The old split has 1016 IDs, which become 1013 after merging by official aliases; cross-split components are 0, with details in summary. The README notes that counts do not prove visual independence, intra-candidate-group ambiguity, or follow-up benefit. For reproduction, code is in the Git repository, inventory source version is 83c8091, and release source version is 36db885c1e785733f43592d4a6eef81711e253b3. It can be run with Python 3.10 standard library using python code/audit_kilogram_inventory.py --config code/configs/kilogram_inventory.json --output <persistent-output> --code-revision 83c8091. Upstream JSON is read only in memory, and output is a field whitelist projection; the complete original files are not in the snapshot. SHA256 values of fixed sources and new outputs are in summary, and all released files are verified by downloading back at this immutable HF revision. This round has no image download, VLM calls, learning-strategy evaluation, or RL training. It is intended to study matching where multiple targets share candidates and are constrained by query costs; this data itself remains static annotations, not a counterfactual answer environment.
SetGround KiloGram元数据
基本信息
- 数据集页面:https://huggingface.co/datasets/gavinlaw/setground-kilogram-tasks
- 许可证:mit
- 语言:en
- 任务类别:other
范围说明
- 只包含KiloGram自然whole/part标注的无身份字段投影、来源校验和划分盘点。
- 不包含图像、参与者ID、时间或操作轨迹。
- 不包含本项目训练模型或交互收益结果。
来源与许可
- 上游:https://github.com/lil-lab/kilogram/tree/f2cd80201eedc5f4882bd88ba32de39195d9e9f3
- 论文:https://arxiv.org/abs/2211.16492
- 上游README明确:除Tangram图像外采用MIT;版权及许可全文随SOURCE_README.md保留。
- 图像来自书籍扫描件,另受教育科研用途说明约束,不能将本仓库MIT标签套用于图像。
文件与Schema
- 快照:snapshots/2026-10-24-source-inventory-v1
- annotations.jsonl.gz:逐行字段source(full/dense)、shape、annotation_index、whole、part(七个polygon ID到自然部件名)。
- 两个来源分别保留,full和dense内重复标注不合并,不应简单相加当独立样本。
- splits.json:保存官方旧ID划分与官方documented_aliases;不提供重新划分。
- summary.json:保存来源blob/SHA256、完整计数、冲突、成本与源码绑定。
- config.json:固定五个源文件。
- SOURCE_README.md和DATASET_README.md:原版许可/数据说明。
计数与划分
- 当前full:1013形状/10173标注。
- 当前dense:74形状/3971标注。
- 旧split:有1016 ID,按官方别名合并后1013个。
- 跨split组件:0个,具体列表见summary。
- 计数不证明视觉独立性、候选组内歧义或追问收益。
复现
- Git:https://github.com/luojiaxuan/cvpr-multimodal-evidence
- inventory源码:83c8091
- 发布源码:36db885c1e785733f43592d4a6eef81711e253b3
- Python3.10标准库:
python code/audit_kilogram_inventory.py --config code/configs/kilogram_inventory.json --output <persistent-output> --code-revision 83c8091 - 上游JSON只在内存中读取,输出是字段白名单投影;完整原始文件不在快照。
- 固定来源与新输出的SHA256见summary,所有发布文件按此次不可变HF revision回下载校验。
- 本轮无图像下载、VLM调用、学习策略评测或RL训练。
- 拟研究多个目标共享候选且受询问成本约束的匹配;本数据本身仍是静态标注,而不是反事实回答环境。





