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.
Mirror of X-Humanoid/RoboMIND2.0-UR5 from ModelScope. All credit to the original authors. Not an official release from them. License unchanged (apache-2.0). For dataset file metadata and data files, please visit the "Files and versions" tab. The current dataset card uses the default template, and the dataset contributor has not provided a more detailed dataset introduction. However, you can download the dataset using the following GIT Clone command or through the ModelScope SDK.
RoboMME-10x is a LeRobot v2.1 dataset containing 16,000 episodes across 16 RoboMME tasks. It contains 7,636,924 frames sampled at 10 FPS. Images are stored directly in the per-episode Parquet files. Dataset metadata is available under meta/, and episode data is arranged as
Video release for RobotEQ. Access is gated: you must request permission and wait for manual approval before downloading any files. This repository is public so the dataset page can be found and cited, but video files are not downloadable until the authors accept your request. You must be logged in (hf auth login) with an account that has been approved. CC BY-NC 4.0. Non-commercial research use only.
A dataset of 7,721 synthetic side-scan sonar images of two target classes, ship and plane, generated with S3Simulator and described in the paper cited below. It is intended as a benchmark for underwater image analysis when labeled real sonar data is scarce. The mine class is released separately in a companion dataset generated with S3Simulator+. intensity (the three channels are identical apart from JPEG rounding, with a maximum difference of 2 levels out of 255). Load as one channel, for example Image.open(path).convert("L"). - metadata/filenamemap.csv maps each released file name to its original file name. - checksums.sha256 holds a SHA-256 hash for every image. Fixed splits are provided…
Training and test data for the Maze experiments in Scale-Equalized Rollout Allocation (SERA).
CrowdMind/SFT-Combinded-USE is a convenience collection of supervised fine-tuning, tool-use, coding, and agent-trace datasets. Source subsets remain separate configurations: no rows are merged, transformed, or forced into a shared schema. Configuration names are short and readable. other, tooluse, and traces identify the source collection; the remaining terms identify the upstream dataset and split. Inspect each configuration's features before use. Their schemas intentionally differ. This repository does not claim authorship of the upstream data or create new annotations. Licenses, attribution requirements, access restrictions, and redistribution permissions vary by source. Users must…
User requested v2 only. Historical v1 score93/150 is retained as a non-matched reference: it preceded the Docker CPU environment fix. New v2 runs use exactly the same OPD149 weights and model budgets with explicit history reasoning removal and CPU-quota-aware sandbox environments. Accuracy, M2.7 measured uncached input/cache reads/output and Docker-separated latency are primary outcomes. No new v1 full150 is required. Two repeats are screening evidence, not proof of non-inferiority.
Action-conditioned safety-judgment data from ManiGuard rollouts across six task families. Inputs: adjacent overview/wrist images, current robot state, the masked unexecuted action suffix, and a natural-language constraint. Targets indicate a new constraint violation during that suffix; monitor outputs are supervision only. The delivery contains raw/ episode records and prepared package/ indices. Training caches are built on the training machine. See SafetyJev training instructions. Preparation in progress. Use the release only when DATASETSUMMARY.json reports complete. Access for external users requires manual approval.
Satellite scenes accompanying SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery. This dataset provides 59 satellite GeoTIFF scenes, totaling 64.79 GB (60.34 GiB). Use these scenes with the SatNav codebase and to run navigation episodes, generate training trajectories, and evaluate policies. SatNav's simulator, SatSim, renders RGB observations from the GeoTIFF scenes as the agent moves through geographic space. Install SatNav following the documentation, then download the episodes and scene files. Set SATNAVSCENESDIR to the local scenes/ directory as shown below. The scene filenames match the scene identifiers used by the episodes. 1. Sign in…
RegalFire — AI Data Foundry Every record has at least two candidate answers. Community votes and accepted status are signals, not scientific truth. Records: 73; distinct source threads: 73; unique answers represented: 186. Domain thread counts: {"statistics": 27, "computationalscience": 25, "biology": 21}. Each record includes a canonical thread, all source answers, author attribution, original HTML, mechanically converted text, revisions and raw hash/API provenance. recordid is the primary key; threadid is site-qualified, while original numeric question/answer IDs are retained. Citation record counts count edges; answer/thread totals are deduplicated within the product. Official Stack…
RegalFire — AI Data Foundry Literal external URLs and DOI-like references with mechanically extracted text context and complete forum attribution. Records: 888; distinct source threads: 221; unique answers represented: 320. Domain thread counts: {"statistics": 61, "computationalscience": 71, "biology": 89}. Actual record splits: {"train": 543, "test": 109, "validation": 109, "holdout": 127}. Each record includes a canonical thread, all source answers, author attribution, original HTML, mechanically converted text, revisions and raw hash/API provenance. recordid is the primary key; threadid is site-qualified, while original numeric question/answer IDs are retained. Citation record counts…
RegalFire — AI Data Foundry Scientific forum QA containing mechanically extracted preformatted code. Code is retained exactly after HTML entity decoding; no execution is claimed. Records: 72; distinct source threads: 72; unique answers represented: 99. Domain thread counts: {"statistics": 34, "computationalscience": 37, "biology": 1}. Each record includes a canonical thread, all source answers, author attribution, original HTML, mechanically converted text, revisions and raw hash/API provenance. recordid is the primary key; threadid is site-qualified, while original numeric question/answer IDs are retained. Citation record counts count edges; answer/thread totals are deduplicated within the…
RegalFire — AI Data Foundry Cases selected by measurable properties: multiple answers/references, no accepted answer, score discordance/spread and answer age span. No correct-answer labels are invented. Records: 258; distinct source threads: 258; unique answers represented: 371. Domain thread counts: {"statistics": 82, "computationalscience": 85, "biology": 91}. Actual record splits: {"train": 161, "test": 32, "validation": 28, "holdout": 37}. Each record includes a canonical thread, all source answers, author attribution, original HTML, mechanically converted text, revisions and raw hash/API provenance. recordid is the primary key; threadid is site-qualified, while original numeric…
RegalFire — AI Data Foundry 89 compact source-grounded statistical QA records. Accepted answers are distinguished from community-score fallbacks. Records: 89; distinct source threads: 89; unique answers represented: 130. Each record includes a canonical thread, all source answers, author attribution, original HTML, mechanically converted text, revisions and raw hash/API provenance. recordid is the primary key; threadid is site-qualified, while original numeric question/answer IDs are retained. Citation record counts count edges; answer/thread totals are deduplicated within the product. Official Stack Exchange API 2.3 source snapshots from Cross Validated (stats), Computational Science…
English · 中文 Real-robot data collected with SONIC for WB-WAM task post-training. LeRobot v3.0, 20 Hz, RGB 360 × 270. The release contains 1,011 episodes, 242,668 frames, 8 tasks, and 12 independent recordings (3.370 hours). Each /recordXXXX/ is an independent LeRobot dataset. Recordings of the same task remain separate: The Dataset Viewer plays the MP4 files. The corresponding Parquet state/action rows and metadata are available under each recording in Files and versions. The action contains absolute next-frame joint targets (+50 ms), not joint increments. Root[3] denotes roll, pitch, and body-frame z angular velocity, not XYZ position. The original train/validation labels are retained in…
Published dictionaries have recorded the Sicilian language in a consistent orthography for several hundred years. Most notably: Pasqualino (1785), Traina (1868), Mortillaro (1876) and Nicotra (1883). Nonetheless, the Sicilian language text that Facebook collected during the No Language Left Behind project (2022) exhibits extreme orthographic diversity which (when incorporated into a training dataset) causes a model to generate a patchwork of different Sicilian dialects with each phrase seemingly sourced from a different town or province. And to make matters worse, very few of the Sicilian-English pairs that the NLLB project collected are translations, which makes the collection almost…
Published dictionaries have recorded the Sicilian language in a consistent orthography for several hundred years. Most notably: Pasqualino (1785), Traina (1868), Mortillaro (1876) and Nicotra (1883). Nonetheless, the Sicilian language text that Facebook collected during the No Language Left Behind project (2022) exhibits extreme orthographic diversity which (when incorporated into a training dataset) causes a model to generate a patchwork of different Sicilian dialects with each phrase seemingly sourced from a different town or province. And to make matters worse, very few of the Sicilian-English pairs that the NLLB project collected are translations, which makes the collection almost…
Published dictionaries have recorded the Sicilian language in a consistent orthography for several hundred years. Most notably: Pasqualino (1785), Traina (1868), Mortillaro (1876) and Nicotra (1883). Nonetheless, the Sicilian language text that Facebook collected during the No Language Left Behind project (2022) exhibits extreme orthographic diversity which (when incorporated into a training dataset) causes a model to generate a patchwork of different Sicilian dialects with each phrase seemingly sourced from a different town or province. And to make matters worse, very few of the Sicilian-English pairs that the NLLB project collected are translations, which makes the collection almost…
Published dictionaries have recorded the Sicilian language in a consistent orthography for several hundred years. Most notably: Pasqualino (1785), Traina (1868), Mortillaro (1876) and Nicotra (1883). Nonetheless, the Sicilian language text that Facebook collected during the No Language Left Behind project (2022) exhibits extreme orthographic diversity which (when incorporated into a training dataset) causes a model to generate a patchwork of different Sicilian dialects with each phrase seemingly sourced from a different town or province. And to make matters worse, very few of the Sicilian-English pairs that the NLLB project collected are translations, which makes the collection almost…
Original RGB-D, UMI, WAX/ADC36, AprilTag, motor/robot telemetry where recorded, dataset metadata, and experiment/training results. Publication was explicitly authorized by the dataset owner. This release is still uploading; catalog.json inventories planned archives, while a completed indices/.json identifies a recoverable archive. Missing signals are not synthesized. Processed RGB-D training samples and their native-label masks are in the companion SightTouch-RGBD-Tactile. Raw records include failed, unknown, evaluation-only, and qualified training sessions; archive integrity does not imply experiment validity, physical calibration, or model accuracy. WAX ADC values remain native codes or…
Small artifacts from the Two-expert transfer experiment: experts E (CruxEval) and E2 (MMLU-Pro) into student B by mixed SFT, mixed RS-SFT, MOPD and weight merges. Experiment report: https://github.com/marin-community/marin/issues/9984. These files were moved from CoreWeave S3 (s3://marin-us-east-02a/marin/) when the experiment's S3 data was cleaned up (marin task #210). Each file keeps its S3 path relative to marin/, so s3origin = s3://marin-us-east-02a/marin/. The S3 copies are deleted after this upload is verified. Totals: 25,735 files, 316.1 MB. MANIFEST.tsv lists every file with bytes, sha256, S3 origin, issue and purpose. Apache-2.0. Teacher-model outputs remain subject to their…
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.