AI Training Datasets

ASTime, short for Activity State Transition Timestamp, is a video dataset for studying real-time perception of human activity changes from a robotic perspective. Instead of only providing the start and end time of an activity segment, ASTime explicitly annotates the transition timestamp: the moment at which a new human activity state becomes semantically distinguishable from the previous state. The released dataset contains anonymized, muted, untrimmed long videos collected in constructed home-like laboratory environments, together with JSON annotations for activity-state transition timestamp localization. The train/test split is subject-disjoint to evaluate generalization to unseen…

TPBench measures three retrieval targets after dialogue compression: This repository includes data, construction inputs, compression and reader code, scorers, saved answers, and paper results. RESULTSMAP.md connects the paper comparisons to their evidence files. Use Python 3.10 or newer. No model access is needed for these checks. The checks validate file hashes and metadata, regenerate all 42 main probe cells, rescore saved answers, and test FMTS context construction. Outputs from tests are written to a temporary directory, not over the released results. Each row fixes the example identifier, method, budget, reference answer, and reader prompts. Identifiers remain stable across…

Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. (grounding>=4, correctness>=3, naturalness>=3). Companion repos hold quality-gate rejects (rejected-) and records whose judge call failed (unjudged-).

The static inventory contains 3,198 YouTube IDs. Earlier collectors stored raw sources under videos/ and records under records/. V5 preserves those files and prepares model-ready frame packs under a versioned prepared/ namespace. New network acquisition samples video streams directly instead of retaining full video files. Existing local sources are reused. Prepared samples contain losslessly compressed RGB uint8 frame banks at 224×224, view indices for micro/mid/long/extralong/spatial, source frame positions, and provenance. Their loader materializes float32 [T,3,224,224] views in [0,1] for the bundled visual model, which applies its own CLIP normalization. Frame counts are 32/16/8/16/12.…

MDS-Bench data Raw data/ folders for MDS-Bench. Layout seconddataset/ /data/... thirddataset/ /data/...

Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. Records judged BELOW the quality gate; each has rejectreasons. Useful for judge calibration, hard-negative mining, or re-filtering with different thresholds. Do NOT use as-is for instruction tuning.

A file is one download, as the source publishes it; an archive counts once. Training use and Redistribution summarize each license as yes, conditional, no or unknown (mixed when the files of a source differ); docs/licenses.md defines each tier and where the license of a file is written. This is a summary, not legal advice: check the upstream terms before use. Browse, search and filter every row in the viewer at the top of this page, or load the catalog: Then download each uri and check it against sizebytes and the hashes (see docs/retrieval.md). A file of several kinds is counted under each of them. The content column of the catalog holds the category. Extensions are those of the downloads…

Dataset · Translation

zh-vie_ecom

Nguyen Thanh Quang

Dữ liệu sản phẩm song ngữ Trung–Việt từ 1688.com, phục vụ tiểu luận chuyên ngành "Tối ưu hoá mô hình dịch máy Việt–Trung cho TMĐT xuyên biên giới". data/snapshot/ là bảng sạch, cập nhật định kỳ — xem ở đây: - bilingualzhvi.parquet: sản phẩm có cả tiếng Trung và tiếng Việt (1688 tự dịch máy). Mỗi dòng có titlezh/titlevi, descriptionzh/descriptionvi (bảng thuộc tính + SKU, đã ghép theo fid nên hai cột song song từng cặp). - monozh.parquet: sản phẩm chỉ có bản tiếng Trung (1688 không dịch, ~43% data/raw/ là log thô append-only (nguồn sự thật, không sửa tay), gồm cả dòng tìm kiếm (chỉ tiêu đề) lẫn dòng chi tiết — không nên dùng trực tiếp để xem hay train, dùng data/snapshot/ thay thế.…

Dataset

211

AI4Manufacturing

[Placeholder] [Placeholder] [Placeholder] [Placeholder] [Placeholder] [Placeholder]

taskcategories: - time-series-forecasting - climate - weather - oceanography - sri-lanka A daily Sri Lankan fishing-harbour weather and marine dataset from 2010-01-01 onward. Data sourced from Open-Meteo (ERA5 and ECMWF).

Dataset · Text generation

flashmini-data-v1

FaithFellowshipFun

Deterministic FlashMini training corpus. Canonical documents live in Parquet+ZSTD shards under shards/; each shard carries a manifest with sha256, counts, and distributions; the frozen corpus identity is corpusfingerprintsha256. recipeonly, gatedrecipeonly, reviewrequired, generatedowned (fail-closed; see registry/sources.yaml + sourcesnapshot.lock.json). Content shards are published only for mirrorallowed/generatedowned sources; recipe-only sources contribute provenance only and stream from upstream at training time. Gated NVIDIA sources are not mirrored. Licenses are recorded per source; no blanket license is claimed over mixed content. filter/dedupe versions, split salt, per-shard…

EventSTVG evaluates event localization and the spatial grounding of semantic roles. This release contains the frozen SoccerSTVG component: 3,516 distinct target events from 50 soccer matches, with 17 semantic event types, 16 semantic roles, and 6,497 event–role instances. All five configurations contain the same 3,516 target event IDs, presented with different video context. They are evaluation configurations, not separate training/validation partitions. Each VideoFolder record integrates its video, complete official prompt, and annotations. No join against an external prompts file is required. The source prompt text, GT answer objects, sample identifiers, and original metadata are…

Every record of the enumeration of four-dimensional 𝒩=1 theories obtained from seed gauge theories with a simple gauge group by adding superpotential terms one at a time, with the superconformal R-symmetry fixed by a-maximization, decoupled operators removed by flip fields, and each candidate tested with its superconformal index (arXiv:2408.02953, arXiv:2510.19136). The records were computed again from the seeds with the package landscape (package 0.6.0 at commit 8046126 of that repository, native extension 0.12.0; pipelinecommit records the development commit a4b6e37 of the same code), level by level until a level has no input. This version supersedes the dataset's first one (commit…

Dataset

trust-chain-freshness

Council of AI

Two things in the Council of AI estate go out of date on their own, and this dataset is the receipt that somebody keeps checking them. An OpenTimestamps stamp is created instantly and carries only a pending calendar attestation. Hours later the calendar's commitment lands in a Bitcoin block — but the published.ots file only says so once the completed path is fetched back and the file rewritten. Nothing does that on its own, so proofs sit reading "pending" long after they are attested, and the publisher looks weaker than the evidence it holds. runs/run-.jsonl has one row per published proof: its state before the run, its state after, and the Bitcoin block heights if it is attested.…