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.
Placeholder. The dataset card is in preparation.
Seventy logged misses from a working research collaboration between Wayfinder and Bones, kept in the order they were made, each with the wrong read, what the record actually showed, and what changed. Most of these misses are Bones's, caught by Bones, by Wayfinder, or by a reviewer named Shuttle. The research behind them is a set of superconducting-qubit runs on IBM and Rigetti hardware and the statements built on them: Seven, with every correction published on the trail. For an earlier essay from the same collaboration, see Consent All the Way Down. Most of them are one move: turning an open measurement into a yes/no verdict too early. Calling a result "unexplained" before running the…
SAM 3D Objects 多视角微调（reconviagensam3d 分支）所用的训练与评估数据。 每个卷是一个 tar.zst，内部路径 = 原始绝对路径去掉开头的 /。解压到 / 即把 数据放回与原机完全一致的位置，所有符号链接自动生效，无需 relink： 或用仓库里的 restore.sh（支持按分组还原、边解边删控制磁盘占用）。 注意第二行：物体的 images/ 链接走的是 renders/... 这条路径，而 renders 自身 只是一个指向 extracted 的目录级软链接。渲染文件实际都存放在 extracted/ 下， 所以 hypersimrenders 分组里除了那些文件，还显式收录了 renders 这个软链接本身； 少了它，新机上 renders/... 路径不存在，全部 images 链接都会断。 打包用 tar 不加 -h，保留链接本身；链接目标由对应分组整体收录。若展开链接， 同一张 HyperSim 渲染图会被数千个物体各拷一份，实测膨胀约 11 倍。 两条 run 使用相同数据。以 poserelabelAB 为例： 完整参数见各 run 目录下的 launch.sh。 训练用 --pointmap-source gt（默认）时，深度不是从物体目录读，而是由 scripts/hypersimso/rawdepth.py 按 cameras/.json 里的 camname / frameidx 实时拼出原始 HyperSim 路径： HYPERSIMRAWROOT 默认…
Apache TsFile version of taku-y/record-20260429-2. A SO-101 follower dataset for the task "Grab the toy 2". Each frame records the 6-D joint state and the 6-D commanded action. Task (taskindex 0): "Grab the toy 2". - 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. - action0..5, observationstate0..5 (FIELD) - FLOAT 6-D SO-101 joint action and 6-D observation state. - frameindex, sampleindex (INT64…
1024-area buckets, no upscale, BF16 AE and text encoder, one source annotation.prompt per image. No original images are uploaded. Standard Musubi text caches retain caption metadata. Each shard has independent tensor archives, a SHA-256 file index and verification report. A shard is complete only when its COMPLETE.json exists. Archives contain cache/ paths; extract all archive parts for a shard into one directory. Use cache filenames to join by source SHA-256. Source lists and plaintext caption manifests stay on the cache coordinator. Only verified shards are eligible for upload. Future machines must claim pending shards through the coordinator before writing output. GPU workers stop after…
This repository documents and packages the data pipeline for the resume-parser-model project. The project explores fine-tuning compact language models to turn resume text into structured JSON, using only facts supported by each resume. The folders preserve the stages used to build the dataset: Resume files and their derived text/annotations use matching category folders and resume IDs so the processing trail can be followed across stages. Each JSONL record contains four top-level fields: - resumeid: the stable resume identifier shared across processing stages. - category: the resume's occupation/category label. - resumetext: the extracted text used as model input (X). - targetjson: the…
A deliberately tiny, instant-loading reverberant speech-like test set — the quick sanity check for dereverberation and robust-ASR front-ends. Every item is a (dry, wet) pair: a clean synthetic speech-like source and its reverberant version through a real-synthetic room impulse response (RIR), with the room's ground-truth acoustics attached. sanity run is cheap. Stratification (balanced across T60 × DRR bins): - t60bin: T60low (<0.3 s) 155 · T60mid (0.3–0.8 s) 143 · T60high (0.8–1.5 s) 93 · T60vhigh (≥1.5 s) 109 - drrbin: DRRvlow (<−8 dB) 163 · DRRlow (−8…−2) 157 · DRRmid (−2…6) 91 · DRRhigh (≥6) 89 Dry sources are synthetic speech-LIKE probe signals, not real speech: seeded harmonic…
A runway catwalk (walk, turn, pose, walk back) for the Unitree G1. Made on the SOMA human body, retargeted to the G1 and converted to SONIC's format. - m7raw/: source motion - m7motionlib/: SONIC-ready motion - scripts/: build the training set - controlslist.txt: the 30 BONES-SEED walk/turn/idle clips mixed in during training (names only, since BONES-SEED can't be redistributed) Team ATR · Motion Data by Bones Studio · Policy: https://huggingface.co/Kyry10/rfcp
Derived from kimtaey/robocasamggr00t300 using prehj/rldx1-robocasa-img-gb64-60k-baseline. Read FORMAT.md before training. Targets are explicit per-observation action chunks; the original action column is retained for provenance and is not the speedup target. 7200 episodes; 2073457 source observations; 2066257 usable anchors. No model was trained for this release. See meta/demospeedup.json and per-episode entropy files.
Preparation notes and schema examples for Robotics tasks using Pointcloud Text data. Full source material is intentionally not bundled, so provenance and licensing remain explicit. - dataset.py — loading, cleaning, and split preparation code. - datasetinfos.json — schema and split metadata. - metadatasample.jsonl — small, human-readable records for checking the schema. - README.md — data card and usage notes. Before using the prepared data, verify source licenses, duplicates across splits, missing values, label balance, and modality-specific corruption. Record the source version and every filtering rule so a later run can reproduce the same rows. The repository is suitable for testing the…
A documented Robotics data-preparation workflow for Text Tabular records. The bundled rows demonstrate the schema and validation path rather than pretending to be a full training corpus. - clean.py — loading, cleaning, and split preparation code. - datasetinfos.json — schema and split metadata. - metadatasample.jsonl — small, human-readable records for checking the schema. - README.md — data card and usage notes. Before using the prepared data, verify source licenses, duplicates across splits, missing values, label balance, and modality-specific corruption. Record the source version and every filtering rule so a later run can reproduce the same rows. The repository is suitable for testing…
This dataset contains 300 simulated expert episodes across six RoboTwin tasks, with 15,455 windows for predicting the motion of object and robot end-effector points. It was prepared for the Trace-MoT experiments. The data is simulation-only. The object queries use simulator segmentation to identify the intended object; this provides oracle target information. - raw/ /episode /episode.hdf5: head, left, and right camera RGB observations, head depth and segmentation, camera calibration, robot end pose, and joint actions. - raw/ /episode /queries.npz and poses.npz: original point queries and simulator pose data used in label generation. - raw/ /episode /instruction.json: task instructions.…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
Apache TsFile version of justintiensmith/rolloutpi05ReasoningStep076596Counting. A SO follower rollout dataset from a pi0.5 policy on a counting task family. The 32 language instructions ask the robot to move objects (a blue pen, a green pencil sharpener, a yellow block and an airplane toy) into or out of a bowl so that the bowl ends up holding a requested number of objects. Each frame records the 6-D joint state and the 6-D commanded action. Tasks (taskindex 0-31) include instructions such as "Leave exactly one object in the bowl", "Adjust the bowl so that it contains two objects", "Leave exactly three objects in the bowl", "Adjust the bowl so that it contains zero objects", "Move the…
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
This dataset was created using LeRobot.
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.