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.
Updated 2026-10-18 · How the library is built
3,164 datasets, sorted by most downloaded.
The evidence behind lua-agent-lab, from lua-agent. The experiment separates structural tool-call validity from useful tool selection and complete task success. The test split is an inspectable demonstration suite, not a training set or a representative benchmark. Do not treat 30 trials as 30 independent problems: ten prompts are repeated with 0, 1 and 2 trailing spaces, matching the original ablation. No confidence interval or population-level claim is implied. Generated with python -m space.exportdataset --tiny --qwen from the source repository. Each measured row includes its UTC collection time, a SHA-256 digest of the relevant source, exact settings and pinned model revisions.…
Each row is one manga panel: the original black-and-white panel (bw) and one painted full-color version (colored), generated by Qwen-Image 2.1 Turbo with one fixed prompt (no tags, no variants). The paint covers the ink, with strong lighting and shading, while characters and scene stay as drawn. Columns: bw, colored (images), panel (path in the source), source (A = file repo, B = parquet repo), series, chapter, and four quality numbers: grayssim (line art survived), meanchroma (colorfulness), blacklit (solid black that came out bright), whitedark (empty paper that came out dark). No quality filtering has been applied. Copyright and content: the source panels belong to their respective…
This public dataset repository preserves the September 2026 MarinSkyRL OPD/MOPD experiment artifacts. artifacts-complete.tar.zst contains the entire original artifacts/ directory, including source checkouts, their Git metadata, the nested local development environment, data, logs, manifests, reports, and symlinks. SHA-256: efcc781203003f67274e23f05e19c6b47679bd761718517ec6241a0e8b0a0b71. The browsable files below are a release copy of the report and evidence directories with links changed to public URLs; the archive preserves the original bytes. To extract the complete tree: zstd -dc artifacts-complete.tar.zst | tar -xf -. The nested.venv is an archival record of the local environment, not…
High-resolution aerial imagery benchmark for remote sensing road extraction, based on the canonical Massachusetts Roads dataset introduced by Volodymyr Mnih (2013). The Massachusetts Roads Dataset covers diverse urban, suburban, coastal, and forested rural environments throughout the Commonwealth of Massachusetts. Standard benchmark partition for road extraction research: If you use this dataset in your research, please cite
Kho dữ liệu âm thanh đa ngôn ngữ (Tiếng Việt & Tiếng Anh) chuẩn phòng thu SOTA, được tiền xử lý và thẩm định toàn diện (SQUIM noise assessment, trích xuất vector âm sắc ECAPA-TDNN 192-d, căn chỉnh phụ đề JSON/timestamps, Mel-spectrogram & F0 pitch contours) phục vụ tối ưu cho việc huấn luyện và suy luận mô hình Matcha-TTS kết hợp Vocoder Vocos. Toàn bộ dữ liệu âm thanh lớn được đóng gói thành các Shard tiêu chuẩn 2.5 GB – 3.0 GB (.tar.gz) để tối ưu hoá băng thông, tránh lỗi timeout và tương thích hoàn hảo với giới hạn Git LFS của Hugging Face. Số lượng: 36 vector đặc trưng âm sắc cho tất cả các giọng nói. Định dạng:.npy (NumPy Array), Shape (192,), Dtype float32, chuẩn hoá $L2$-norm = 1.0.…
Artefatos do estúdio de geração de vídeo da Maya — a personagem é uma modelo digital adulta, criada e mantida pelo dono do projeto. Este repositório guarda o que não se recria: as receitas, as âncoras aprovadas, o código dos runners e os vídeos entregues. O disco da máquina de trabalho (kvm2) chegou a 89% de uso. Artefato que já cumpriu o papel não pode ficar empilhando na máquina onde o trabalho acontece — daí ele subir para cá, onde fica recuperável e versionado. - codigo/ — o que faz o estúdio funcionar - motor/ — captura de tela do celular por ADB e automação do telefone - receitas/ — as receitas de geração (Klein, LTX, H3) com os parâmetros medidos - imagem-maya/ — os scripts de imagem…
Twitch VODs of MCSR Ranked matches (Minecraft 1.16.1 random-seed any% speedruns, Diamond+ players), collected for a pixels-to-keyboard speedrun agent. Each VOD comes with its inverse-dynamics-model targets. - vods/ /.mp4: the VOD (360p30 as downloaded), sharded by the 2nd and 3rd characters of the name. - vods/ /.idmtargets.npz: per-frame targets aligned to the VOD (from the match replay, see Infatoshi/mcsr-speedrun-train). - raw/: a few full downloads that were not yet processed. Source: the MCSR Ranked API (match index, VOD links) and Twitch. The footage belongs to its streamers; this is a research mirror. Companion repos: Infatoshi/mcsr-speedrun-train (training shards, replays…
Extracted contextual hidden vectors from openbmb/MiniCPM5-2B at Layer 32 (Hidden Dimension: 2048) in bfloat16. Designed for JEPA representation learning.
This package contains the reproducibility record for the Agentic Film Lab video testing H3 Base, Realism People, and Better Human Motion with recurring characters. - manifest.json: the eight-cell test matrix, exact seeds, steps, LoRA strengths, geometry, prompt filenames, workflow filenames, and prompt SHA-256 hashes. - prompts/: the exact text supplied to each generation. - workflows/: the eight ComfyUI API workflows used for the viewer-facing runs. - anchors/: the four starting images used by the matched pairs. - cost-ledger-v001.md: generation time, displayed-rate estimates, provider billing, diagnostic activity, and accounting definitions. - observations.md: the contemporaneous output…
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.