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.
A packed pretraining corpus with 100,000,000,000 training input positions and 10,000,000 held-out validation positions, prepared for the recurrent Purrence Hyperbullet language model. It uses a frozen 2,048-ID tokenizer (2,047 ordinary IDs plus structural boundary ID 2,047). Upload status: consider this repository complete only when UPLOADCOMPLETE.json is present. That receipt identifies the immutable data revision whose remote file sizes and checksums were verified. This is a custom memory-mapped corpus, not a Parquet dataset for datasets.loaddataset. Download the core training files with the Hugging Face Hub client: For reproducibility, pass the revision from UPLOADCOMPLETE.json after…
Apache TsFile version of Shiki42/PutCab-AB-RandomDelta-Train50. A timing-balanced 50-episode subset of RoboTwin PutCab-AB demonstrations, recorded on a simulated ALOHA bimanual robot. Selection is conditioned on the normalized reference offset uref = (delta + TBref) / (TAref + TBref) (where delta is right/B start minus left/A start): ten intervals 0,.1)... [.9,1) each contribute five trajectories, giving 25 negative and 25 positive deltas. Reused trajectories keep independent active-arm v3 timing; the new V2 trajectories add recorded feedback waits (safe-waypoint lift, drawer-readiness re-check). observation.state holds 14 joint/gripper drive targets and action is the next observation…
Preparation notes and schema examples for Geology tasks using Image Depth 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 data…
2,500 train and 250 validation rows. Long examples have 4,096–7,168 thinking tokens; complete ChatML samples including final answer and EOS fit in 8,192 tokens. Long source: open-r1/OpenR1-Math-220k at e4e141ec9dea9f8326f4d347be56105859b2bd68 (only math-verified generations). Short replay source: danghoang2005/qwen25-math-sft-long-15k-v1 at 32f0afc54a09c4963ae2478cf93871025abda901. Tokenizer: Qwen/Qwen2.5-Math-1.5B at 4a83ca6e4526a4f2da3aa259ec36c259f66b2ab2. Deduplicated across splits and screened against GSM8K, SVAMP, MATH-500, and AIME 2026. Review the source datasets' licenses and terms before reuse or redistribution. See manifest.json for counts and reproducibility settings.
Public final records for the Qwen3-1.7B student / Qwen3-4B-Instruct-2507 teacher PG-OPD experiments. All 17 Test1 measurement/comparison sets, 9 masks, 40 training records, 41 primary AIME24 evaluations, 29 AIME25 chat evaluations and 116 plain-protocol benchmark evaluations are included. Evaluation protocols remain separate; see experimentconfig.json and experimentreportopd.md. All 52 checkpoint directories remain in https://huggingface.co/yufanzh/qwen3-opd-experiment-archive without reupload. Download a full runs/ /checkpoint- / directory to load a model. These contain model weights, config and tokenizer, not optimizer-state snapshots. finalfilemanifest.json lists the current records and…
Checkpoint-specific RADGRAM2 calibration statistics for converting the original BF16 Qwen/Qwen3.8-Flash-Next checkpoint with Radiance's GPTQ converter. These are calibration Grams, not a fine-tune, adapter, model checkpoint, or training dataset. They are intended for Radiance conversion recipes that use w4nl/w5nl expert quantization. - grams/part-00/, grams/part-01/, and grams/part-02/: all 24,768 RADGRAM2 files for both tensor-parallel ranks. The files are sharded because the Hub limits a single directory to 10,000 files. - reassemble-grams.sh: creates the flat grams-flat/ directory expected by Radiance. - grams-validation.json: structural and coverage validation results.…
Seven short videos prepared for broad VLM and video-decoder compatibility. The root metadata.jsonl follows the Hugging Face VideoFolder layout. Its filename field points to each file under videos/; the remaining fields record source names, media properties, and SHA-256 checksums. The dataset repository is public and can be loaded without authentication.
This dataset was created using LeRobot.
Continuation of xxxspatialencoderinfinigen2/rawdatainfinigen2: Blender scene files (scene.blend) of the augmented Infinigen indoor batch indoors5000additional (4,080 original scenes x up to 20 augmented layouts, 79,275 files, 18.9 TB), packed as WebDataset tar shards of about 5 GB and grouped into parts of about 2 TB. Each sample has.blend and.metadata.json (original path, size, sha256). Each part directory ends with manifest.json (all shards and their rows) and SHA256SUMS; a part without them is still uploading. The rendered RGB-D frames and OBB annotations for these scenes are in xxxspatialencoderinfinigen3/rawdatainfinigen3 (infinigen5000/).
Auto-generated sample of the labelled boxes (with judge scores when available). Regenerated on every push.
RealEstate10K camera trajectories and frames, in the pixelSplat chunk release, repacked as random-access WebDataset shards. Every file in this repository was verified against the source copy (same file set, same byte sizes). WebDataset tar shards (about 3 GB each), one tar row per scene, indexed for random access with ultar.utix files (msgpack rows of entry keys, offsets and sizes, so they can be read without ultar). Every scene row holds: -.tar.utix — the ultar row index (offsets and sizes of every entry). -.tar.centres.bin — float32 camera centres per frame per scene (a small sidecar used for farthest-point view sampling); header DPFV, then a per-scene offset table in index row order.…
Daily financial time series from 2008-01-01 through the inclusive cutoff 2026-06-30. Status: readywithgaps. The universe is frozen from Abner0803/Trading-Benchmark-raw at 84908ce; historical constituents are retained, including symbols no longer available from providers. raw/stock.pkl: trusted Python dicts of ticker → DataFrame, preserving the published raw column conventions (datetime for CSI, Date for US/Japan). processed/ / /data.parquet: long-form features and labels. quality/.json: per-ticker coverage, retrieval status and exclusions. manifest.json: coverage, provenance, dependency versions and SHA-256 checksums. 2020-01-01 ≤ date < 2021-01-01; test: 2021-01-01 ≤ date ≤ 2026-06-30.…
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.