Single val split: 1,013 retained records from 303 source train + 744 source validation records of tridm/VMLU. Source train: 292 retained / 11 excluded; source validation: 721 retained / 23 excluded. All 58 subjects remain represented. No test data was downloaded or included. Only these nine columns are exported in the Parquet dataset: Every promptaudio begins with: The remaining answer-letter instructions and all original choices are unchanged. For text evaluation use prompt. For audio evaluation supply promptaudio together with audio; do not send question, transcription or ttstext as additional model context. The audio contains the question only, without options, answer keys or model…
SAVRN Model Hub
AI Training Datasets
Each dataset with its card, its structure (every split, row count and column type), its files, its license, and the models that disclose training on it.
11,442Models
3,247Datasets
285Papers
6,976Publishers
12,091Sourced relationships
Updated 2026-10-19 · How the library is built
3,247 datasets, sorted by most downloaded.
Original manuscript photographs from the HMML Vööbus collection, packaged with provenance metadata for OCR/HTR research and transcription. Release status: inprogress. 7,379 image records uploaded out of 75,581 catalogue entries. The catalogue configuration contains the full inventory; the images configuration contains only successfully downloaded, decoded, and uploaded images. See ingestion-status.json for the latest counts and ingestion-errors.jsonl for unresolved records. Each image is embedded in a Parquet shard. Loading images does not depend on HMML's server. JPEG bytes are preserved without resizing or recompression. Dimensions and SHA-256 hashes refer to these original files. The…
Textbook Cover VQA is a Visual Question Answering (VQA) dataset developed for research on structured understanding of text-rich academic textbook covers. The dataset is designed to evaluate vision–language models (VLMs) on textual understanding, academic-domain recognition, spatial grounding, and visual content understanding. Textbook Cover VQA consists of 1,510 textbook-cover images covering 9 academic domains. Each textbook-cover image is associated with 8 predefined VQA questions, resulting in 12,080 question-answer annotations. The dataset captures four major types of information: - B. Vinay Kumar — L24AEC592 - A. Akash — Y23AEC404 - A. Pujitha — Y23AEC410 - B. Avinash — Y23AEC413…
The frozen calibration, probe, and ladder sets underpinning every published number in the warp-3090-quant-lab repo (W4A16 GPTQ quants + DFlash2 speculative drafters for Qwen 3.5-family models). These files are deliberately frozen: the headline metrics are defined on these exact rows — - calib-cascade-v1.jsonl (999 rows) — the champion GPTQ calibration set (weightedKL 0.1087 @ 4B, 0.0889 @ 9B, 0.1264 @ 27B trace to it) - probe-v4.jsonl (378 rows) — the fidelity ruler; the weightedKL metric - probe-guard.jsonl (76 rows) — OOD guard (norobots + Belebele, 9 languages) - ladder-20.jsonl (20 prompts) — the acceptance ladder; the τ metric (shipped drafter τ 3.79 at the serving point / native MTP…
This dataset comprises real RGB images of weeds in agricultural fields at Montoldre, France, specifically within maize and bean crop areas. Captured using a robot platform during 2019 and 2021, the images were collected under natural field conditions to support semantic segmentation research for weed detection in crop environments. The dataset contains 985 images with pixel-level mask annotations. This dataset is indexed on https://project-agml.github.io/ as part of the AgML python library. This dataset was reformatted from its original format to match HuggingFace standards.
WildChat-derived checklist/requirements data, excluding examples already present in yikeee/wildchecklistsopenrubricsrubrics. Rows from combinedwildchatrequirements whose instruction appears in yikeee/wildchecklistsopenrubricsrubrics (matched via messages[0].content) were removed. - Removed (instruction overlap, including duplicate instructions): 62,617 - instruction: user prompt - rawresponse: model/raw checklist generation output - requirements: parsed requirements list - origrow: original WildChat row metadata - origrowidx: index in source file - sourcefile: source file path
This repository mirrors archives of two WILDS datasets for download and staging on compute nodes. It is not an official WILDS repository. - camelyon17v1.0zip/: six independent ZIP files. Each contains paths under camelyon17v1.0/. - povertyv1.1zip/: seven independent ZIP files. Each contains paths under povertyv1.1/. Each ZIP can be extracted separately. Extract all archives into the same root directory to reconstruct both datasets: The resulting directories are wilds/camelyon17v1.0/ and wilds/povertyv1.1/. Pass wilds as the rootdir when loading these datasets with WILDS. See the WILDS dataset descriptions and cite the original datasets in research using them. WILDS describes Camelyon17 as…
One folder per day under data/YYYY-MM-DD/, and the same files copied to latest/. Built every morning (JST) by the EmpireOS world model. Nothing is generated by a model; every row is a measurement from a public source. Countries: JP, US, GB, DE, FR, ES, IT, BR, MX, IN, ID, KR, TW, VN, TH, PH, TR, SA, EG, NG, ZA, KE, PL, NL, SE, AU, CA, AR, CO, CL. Agents and analysts keep re-scraping the same public endpoints every day. This dataset does it once, keeps the history, and gives it in one flat shape. GDELT data is free for any use with attribution to the GDELT Project. Google Trends, Apple, GitHub, Hugging Face, Steam and npm data are reproduced as published by their public endpoints; check each…
This public dataset contains 65,131 rendered counterfactual cases organized into 8,491 same-history action/future groups. Each case has 4 historical and 8 future CAMF0 frames plus WorldEngine metadata. Archives preserve complete pair groups. The audit bundle contains group JSON manifests, receipts, validation reports and the completion contract. Large generator-intermediate subset PKLs are intentionally excluded because the CAST paired-cache builder consumes the group JSON and rendered frames only. Download all files and run bash extractall.sh DESTINATION. The extraction script verifies every archive against SHA256SUMS first.
WorldPop's gridded population of Japan, 2015 to 2030, release R2025A: how many people live in every 100 m and every 1 km cell, by sex and five-year age band at 1 km, and the degree of urbanisation of every cell. 1,024 rasters as WorldPop published them, and the same rasters as Cloud Optimized GeoTIFF. CC BY 4.0. Cite each product by its DOI (see Licence). WorldPop's server sends whole files and ignores HTTP Range requests, at about 400 KB/s a connection, and its GeoTIFFs are striped. To look at one city you fetch the country. The COGs here are tiled at 512 with overviews, and the Hub answers Range requests, so GDAL's /vsicurl/ or rasterio can read a window. The files of a release are not…
Apache TsFile version of zwonzis/xarm7-pick-orange-ball-large. A UFactory xArm7 dataset for the task "pick up the orange ball and put it in the cyan bowl". Each frame records the 8-D arm state and the 8-D commanded action. Task (taskindex 0): "pick up the orange ball and put it in the cyan bowl". - Time (INT64, milliseconds) - per-episode sample time, round(timestamp 1000) in milliseconds (the source timestamp column equals Time / 1000 seconds and is therefore not stored separately). - episodeindex, taskindex (TAG) - device dimensions identifying one demonstration episode and one task; query a single episode with WHERE episodeindex=0. - observationstate0..7, action0..7 (FIELD) - FLOAT 8-D…
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
Embedding models for retrieval
Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher.
Collection · 4 entries
Models that fit on one accelerator
Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer.
Collection · 6 entries
Open-weight text models worth knowing
Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
Related SAVRN Research
The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them.
SAVRN Index
What open models cost to run
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.