scaling-literatures-cuabench-djaiiofjhaio
收藏资源简介:
该数据集名为“全量种子与文献素材”,面向科学类种子题构题产线中的素材预处理,来源为 SciAILab/OSWorld-Science-data。它覆盖 8 个领域、153 道种子题,并排除 VM 环境相关内容。其用途是采集、整理和同步种子题及其配套文献素材,而不是直接生成或验收新题。README 明确指出,本库不代表已生成或验收 153 道新题;本批也仅进行素材采集和入库,不构题、不运行 debate loop,也不宣称已完成科学实验验证。数据内容与构成包括:标准化种子及原始附件,保存在 inputs/seeds/;按 DOI 保存的文献正文、补充材料 SI、关联 GitHub/HF 材料及来源回执,保存在 inputs/papers/;候选 DOI 与种子关联文件 candidates.json;待获取正文清单 paper_waitlist.json;实时逐题汇总 paper_seed_materials_summary.md。同步到 HF 的范围包括 inputs/、README、candidates.json、paper_waitlist.json 和 paper_seed_materials_summary.md;空目录跳过,仅上传变化文件,远端不自动删除;凭据、运行日志、缓存与安装环境不上传;同步记录保存在本地 maintenance/hf-sync/ 中,定时启动不等于上一轮已上传成功,应以远端文件校验为准。本地工作目录中的 dataset/、paper_seed_materials_summary.json、scale_state.json、runs/、imports/、upstream-metadata.json 及维护记录不属于 HF 同步范围。HF 上的汇总 Markdown 是最近上传快照,并非本地实时状态。截至 2026-10-31T13:09:11.028311+08:00 的进度快照显示:种子题及附件本地采集 153 道,HF 远端已核验 153 道,全部文件哈希一致;main 与 SI 齐全文献本地采集 368 篇,HF 远端已核验 190 篇;达到各题文献目标的种子题本地为 132 道,HF 远端为 66 道。README 同时提到当前核对版本的 HF 素材库共 5,900 个文件,约 14.31 GB,版本为 e72506b2336b244ed7636b6a017b9e067b6b066d。远端完整文献统计以本地清单为基准,逐个核对正文与 SI 的内容哈希;仅有正文、仅有 SI、缺失或哈希不同均不计入。题目达标数按各题候选 DOI 与各自目标计算,同一文献可服务多道题。当前仍有 178 篇本地完整文献尚未在远端核验齐全,因此不能把本地采集数当作已上传数。采集与同步流程通过 pipeline scale-papers 入口执行:固定上游版本并核验全部非 VM 输入;每题导入后,由独立文献采集 CLI 按标题、摘要与 capability 筛选,再下载正文、SI 及关联 GitHub/HF 材料;每轮 10 篇新候选,整轮结束后按各题目标检查完整文献数量,默认目标为 3 篇,显式配置可不同,超额全部保留,同一 DOI 的物理文件不重复下载。自动同步每 10 分钟触发,若上一轮仍运行则不重入;传输、网络或提交限额可能导致远端滞后。该数据集适用于种子题素材预处理、文献采集与同步管理、候选文献与种子关联维护、采集进度核对等场景;不适用于宣称已完成题目生成、验收或科学实验验证的场景。
The dataset, named Full Seed and Literature Materials, is intended for material preprocessing in a scientific seed-question construction pipeline. It is sourced from SciAILab/OSWorld-Science-data. It covers 8 domains and 153 seed questions, excluding VM-environment-related content. Its purpose is to collect, organize, and synchronize seed questions and their accompanying literature materials, rather than directly generating or accepting new questions. The README explicitly states that this repository does not represent 153 new questions that have been generated or accepted; this batch only performs material collection and ingestion, does not construct questions, does not run debate loops, and does not claim completed scientific experiment validation. Data contents include standardized seeds and original attachments in inputs/seeds/; DOI-organized paper main texts, supplementary materials (SI), associated GitHub/HF materials, and source receipts in inputs/papers/; candidates.json for candidate DOI-to-seed associations; paper_waitlist.json for a to-be-acquired main-text list; and paper_seed_materials_summary.md for real-time per-question summaries. HF synchronization includes inputs/, README, candidates.json, paper_waitlist.json, and paper_seed_materials_summary.md; empty directories are skipped, only changed files are uploaded, remote files are not automatically deleted; credentials, run logs, caches, and installation environments are not uploaded; sync records are stored locally in maintenance/hf-sync/. A scheduled start does not mean the previous round uploaded successfully; remote file verification should be used. Local working directories dataset/, paper_seed_materials_summary.json, scale_state.json, runs/, imports/, upstream-metadata.json, and maintenance records are not within HF sync scope. The summary Markdown on HF is the latest upload snapshot, not the local real-time state. As of the snapshot at 2026-10-31T13:09:11.028311+08:00: seed questions and attachments: 153 collected locally, 153 verified remotely on HF, all file hashes consistent; papers with complete main text and SI: 368 collected locally, 190 verified remotely on HF; seed questions meeting their literature targets: 132 locally, 66 remotely on HF. The currently checked HF material repository has 5,900 files, about 14.31 GB, version e72506b2336b244ed7636b6a017b9e067b6b066d. Remote complete-literature statistics are based on the local manifest, verifying main-text and SI content hashes one by one; main-text-only, SI-only, missing, or hash-mismatched items are not counted. Per-question target achievement is calculated from each questions candidate DOIs and target, and the same paper can serve multiple questions. There are still 178 locally complete papers not fully verified remotely, so local collection counts cannot be treated as uploaded counts. Collection and sync process: executed via the pipeline scale-papers entry; fixed upstream version and verification of all non-VM inputs; after each question is imported, an independent literature collection CLI filters by title, abstract, and capability, then downloads main texts, SI, and associated GitHub/HF materials; 10 new candidates per round; after each round, complete literature counts are checked against each questions target, default target 3, explicitly configurable, excess retained, and physical files for the same DOI are not downloaded repeatedly. Automatic synchronization triggers every 10 minutes; if the previous round is still running, no re-entry occurs; transfer, network, or submission limits may cause remote lag. Suitable for seed-question material preprocessing, literature collection and synchronization management, candidate literature-to-seed association maintenance, and collection progress checking; not suitable for scenarios claiming completed question generation, acceptance, or scientific experiment validation.
全量种子与文献素材
基本信息
- 数据集地址:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio
- 来源:https://huggingface.co/datasets/SciAILab/OSWorld-Science-data
- 覆盖范围:8 个领域、153 道种子题,排除 VM。
- 用途:用于构题产线的素材预处理,不代表已生成或验收 153 道新题。
当前采集与上传进度
- 核对时间:2026-10-31T13:09:11.028311+08:00。
- 说明:这是一次进度快照;持续采集的最新状态见逐题汇总:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/paper_seed_materials_summary.md
| 指标 | 本地采集 | HF 远端已核验 |
|---|---|---|
| 种子题及附件 | 153 道 | 153 道,全部文件哈希一致 |
| main+SI 齐全文献 | 368 篇 | 190 篇 |
| 达到各题文献目标的种子题 | 132 道 | 66 道 |
HF 素材库核验状态
- 当前核对版本:e72506b2336b244ed7636b6a017b9e067b6b066d。
- 文件数量:共 5,900 个文件。
- 文件大小:约 14.31 GB。
- 远端完整文献统计以本次本地清单为基准,逐个核对正文与 SI 的内容哈希;仅有正文、仅有 SI、缺失或哈希不同均不计入。
- 题目达标数按各题候选 DOI 与各自目标计算,同一文献可服务多道题。
- 当前仍有 178 篇本地完整文献尚未在远端核验齐全,不能把本地采集数当作已上传数。
- 自动同步每 10 分钟触发,若上一轮仍运行则不重入;传输、网络或提交限额会使远端滞后。
采集与同步流程
- 入口为 pipeline scale-papers。
- 固定上游版本并核验全部非 VM 输入,每题导入后由独立文献采集 CLI 按标题、摘要与 capability 筛选,再下载正文、SI 和关联 GitHub/HF 材料。
- 每轮 10 篇新候选,整轮结束后按各题目标检查完整文献数量;默认目标为 3 篇,显式配置可不同,超额全部保留。
- 同一 DOI 的物理文件不重复下载。
- 同步范围:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/tree/main/inputs 、数据集页面 README、https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/candidates.json 、https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/paper_waitlist.json 、https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/paper_seed_materials_summary.md 。
- 空目录跳过,仅上传变化文件,远端不自动删除。
- 凭据、运行日志、缓存与安装环境不上传。
- 同步记录在本地维护目录中的 hf-sync 子目录;定时启动不等于上一轮已上传成功,应以远端文件校验为准。
素材目录与说明
- 实时逐题汇总:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/paper_seed_materials_summary.md ,包含采集进度、完整篇数与逐题目标。
- 候选 DOI 与种子关联:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/candidates.json 。
- 待获取正文清单:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/blob/main/paper_waitlist.json 。
- 标准化种子及原始附件:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/tree/main/inputs/seeds 。
- 按 DOI 保存正文、SI、关联库与来源回执:https://huggingface.co/datasets/StarLiu714/scaling-literatures-cuabench-djaiiofjhaio/tree/main/inputs/papers 。
本地保留且不同步内容
- 固定上游快照、结构化汇总、scale_state.json、runs、imports、upstream-metadata.json 及维护记录仅保留在本地工作目录,不属于 HF 同步范围。
- HF 上的汇总 Markdown 是最近上传的快照,并非本地实时状态。
范围声明
- 本批仅采集和入库素材,不构题、不运行 debate loop,也不宣称已完成科学实验验证。





