Single-speaker synthetic Tajik speech, generated 2026-09-22..24. Sources: khazina (book readings), law (legal texts), books (book texts). CC-BY-NC-4.0. Research use only. Generated with an OmniVoice LoRA (omnivoice-lora-sl-panther2); OmniVoice is CC-BY-NC, so this corpus inherits the non-commercial restriction. Every manifest row carries licensenote: "CC-BY-NC-derived, research-only, never Sado product". This data must not be used in a commercial product. Manifest fields: clipid, text, wav, duration, refvoice, generator, synthetic, rmsdbfs, normalizationtargetdbfs, peakceilingdbfs, normalizationmethod, licensenote, sourcerow. All audio is loudness-normalised to -16 dBFS RMS with a -1 dBFS…
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.
11,122Models
3,164Datasets
285Papers
6,792Publishers
11,852Sourced relationships
Updated 2026-10-18 · How the library is built
3,164 datasets, sorted by most downloaded.
This is a synthetic training corpus generated by the Smolify Foundry. It was used to train the corresponding model smolify/smolified-aiexpense. This dataset is a sovereign asset owned by smolify. Generated via Smolify.ai.
Verified recovery SFT trajectories for visual web agents. - 2000 multi-turn conversations in policysft.jsonl - Benchmarks (as even as the accepted pool allowed): MiniWoB 639, TimeWarp 639, WebArena-Verified 362, VisualWebArena 360 Image files referenced by path live on the collection machine under /root/SFT200020260916/batches/. This JSONL is the training text; pair it with those PNGs if you train a vision model.
A dataset of news about snowfall, snowstorms and blizzards from a global news index, January 2019 – July 2026, built with the Prism News API (historical and real-time news, https://radar-api.fliarbi.com/news). Every row is one news article about snowfall, with its date, location, tone, editorial priority, themes and the organisations it names, plus a link that returns the full article through the API. - 59.76% of each year's snowfall coverage falls in December–February (2019–2025 average); June–August holds 5.57%. - Snowfall's share of news ranged from 4.08 to 7.97 articles per 10,000 between 2019 and 2025; the 2025 level was 39.4% below 2019. - January 2026 (the January 23–27 North…
This dataset was created using LeRobot.
这是一组公开的SO-101实机诊断录制，用于复核SmolVLA在“夹住抽屉边缘→松开→自主返回初始区域”任务中的正常表现与失败。当前只有2026-10-16的5条SmolVLA录制，1条用户反馈正常、4条用户确认失败，不是正式留出测试集，也不是ACT/DP/SmolVLA成功率排行榜。 五条共4282帧，top/wrist视频均640×480，名义30FPS。原始data/meta/videos保持完整；单条30秒录制中的重复动作周期不是多次独立试验。 - 每条目录的labels.json是附加标签，原始LeRobot文件没有被改写。 - 未明确确认的人工介入、重试次数、实际控制频率使用unknown/null，不根据视频背景或夹爪变化次数补猜。 - 这些录制已经参与问题分析，不能再作为未经接触的最终留出测试集。meta/info.json中的默认train split只保留录制格式，不代表训练用途。 模型为SmolVLA 15K，HF根目录本身是20K，不能混淆。实际使用权重SHA256为7a02885f974b22d50959c8d7e9cda94f5ab2c861af959a25391e814821449f85。 配置/提供的CLI为chunksize50、nactionsteps50、nobssteps1、Flow numsteps10，CUDA同步推理；日志没有完整配置转储。实际模型输入top→camera1、wrist→camera2，六维状态/动作。自动复位禁用。 本仓库根目录是评测合集，不是一个合并后的LeRobot根目录。以下下载轻微改写条件：…
This dataset was created using LeRobot.
This dataset was created using LeRobot. SO-101 turntable pick-and-place (robot lilguy), September 2026. A single SO-101 follower arm, teleoperated from an SO-101 leader arm, picks a cube off a slowly spinning turntable and places it in a basket. Two cube colours (blue, green) and two operator speeds (slow, fast). Recorded with MakerMods Lab on 2026-09-27 and 2026-09-29. - observation.state: 6 follower joint positions - action: 6 leader joint positions - observation.images.head: Intel RealSense D435if RGB, 640×480, overhead view - observation.images.headdepth: RealSense depth aligned to the head RGB, 12-bit lossless HEVC (decode with lerobot's depth support to get millimetres; values under…
This dataset was created using LeRobot.
Action100M-inspired annotation artifacts using V-JEPA 2.1, configurable captioners (default Qwen2.5-VL video and Qwen3-VL image), and optional Claude aggregation. This is an adaptation, not a reproduction. Exact models are recorded in run.json. Each train row represents one captioned or requested aggregation node. Context-only nodes remain in the episode sidecars and are omitted from aggregation rows. selectedsubtask marks the controller cut; source frame intervals are half-open [startframe, endframe). The split name does not imply a held-out evaluation split or verified labels. Source dataset IDs/revisions and episode IDs identify the original demonstrations. captioning/ contains caption…
This dataset was created using LeRobot.
Apache TsFile version of nikodembartnik/so101pencilholder1. A SO follower dataset for the task "put the pen in the wooden box" (source pretty name: "Task: put pen in wooden box"). Each frame records the 6-D joint state and the 6-D commanded action. Task (taskindex 0): "put the pen in the wooden box". - 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…
Teleoperated demonstration dataset for training a grasping policy on an SO-ARM101 arm with an AmazingHand dexterous hand (imitation learning). Pick up the cube with the dexterous hand — grasp a cube on the table using the dexterous hand. Collection flow: an operator moves the leader arm → the follower tracks it while its gripper proportionally drives the hand's open/close → joint angles and both camera streams are recorded synchronously. Per-episode flow: start from a fixed initial pose → perform one complete grasp (approach → grasp → lift → move → place) → end. Position coverage: the cube was placed at 4 different table positions, 5 episodes each, to teach positional generalization.…
Preparation notes and schema examples for Social tasks using Audio Text data. Full source material is intentionally not bundled, so provenance and licensing remain explicit. - builddataset.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…
Spanish from 12 speakers born in 9 countries. 19 clips, 0.3 hours, each labelled with the speaker's country of birth, first language, self-reported accent or regional variety, declared Spanish proficiency, age band, gender and recording device. Audio and speaker metadata only. Human-validated transcription is available on request. Silencio's Spanish catalogue is 65% Latin American by recorded hours, led by Venezuela, with Spain at 9%. It also holds a substantial body of Spanish spoken as a second language: 18% of hours come from speakers born in Africa, notably Nigeria, Kenya and Egypt. This sample covers 9 countries of birth across Latin America, Spain, Africa, the Middle East and the…
Finished MIDI-guided music edits made with SpanSynth-Edit. Open a work's shared link to compare the original and edited audio, explore its score, and make an editable copy. Each work has its own folder: - input.wav and output.wav: the original clip and saved generated audio, at 48 kHz mono. - original.mid and edited.mid: the source and edited score, aligned to the clip. - context.wav: the normalized audio used by the model, including the history before the clip. - project.json: the title, description, notes, editing region, generation settings, and piano-roll display settings. Shared links have the form https://mimbres-spansynth-edit.hf.space/?work=FOLDERNAME. An unlisted work is still…
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.