l4096Qwen3-1.7Bcompressionbs32n1632kt1noeosverl091 One verified gzip JSONL shard per training step; 512 responses per shard. Historical compression MathVerify after thinking, without an EOS gate.
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.
Synthetic data only. Transcripts, word error rates by noise level, the draft notes and the sign-off checklist from one run of MIS 752 Lab 12.
Version française plus bas. Synthetic data for the BuildersLab trial period: short starter assignments for new cohort members. Lagune Market is a made-up online marketplace based in Abidjan, Côte d'Ivoire, with prices in FCFA and payments by mobile money, card or cash on delivery. Every customer, order and document is invented. Generated with a fixed seed by scripts/trial-datasets/ecommerceorders.py in the BuildersLab repository. The files come from a French-format Excel export, so they need some care when loading. The answers are not in this repository. One customer can place many orders, and one product can appear in many orders. Each folder has a README.md with full instructions. Données…
Raw, byte-identical copies of two Los Alamos National Laboratory public cybersecurity datasets, acquired exclusively from the official LANL source (https://csr.lanl.gov/). No mirrors, Kaggle copies, or repackaged distributions were used. - Filenames are preserved exactly as distributed by LANL. Each archive is stored twice: the original compressed file (byte-identical, never re-encoded) and its decompressed form at the same path minus the.gz/.bz2 suffix. - Decompressed files are produced locally with gzip -dc / bzip2 -dc and verified on the repo by SHA-256 (LFS OID or byte comparison). - LANL does not publish netflowday-01.bz2; day-01 netflow is unavailable from the official source.…
This data card accompanies a lightweight Legal loader for Image Text metadata. It is meant for pipeline inspection, source adaptation, and reproducible split preparation. - prepare.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…
LeRobot v2.1 dataset of LIBERO-10 demonstrations perturbed along the end-effector position x axis at scale 0.2 (alpha=0.2), with sequential square-wave injection (half-period = 1).
The rollouts of Donghyun1228/libero-policy-rollouts-success-20261002 (successful episodes) and Donghyun1228/libero-policy-rollouts-pooled-20261002 (all episodes) re-packed as one LeRobot v2.1 dataset per camera view, so a single repoid can feed the policy (imitation) loss. Rows are unchanged; the four suites are concatenated with episode, frame and task indices renumbered. meta/episodes.jsonl keeps source: {suite, episodeindex} back to the per-suite datasets. The IDM windows stay in the source repos' /lerobot/alignedidmh10.json manifests and refer to the same episodes. the policy saw only the wrist camera (external camera masked at inference); stored images are always the shifted external…
50,000 frames across40 tasks at20Hz; each task uses the first1,250 frames of the pinned source. 49,600 H10 windows, with no hidden exclusions. MuJoCo3.2.3. Source actions are physically replayed at yaw0, then the simulator state is transferred to compensated yaw20. The external camera is agentviewmedium; wrist images and7D actions retain their normal schema. Saved translation/rotation commands are rotated by -20degrees; gripper is unchanged. This dataset follows the same state-rerender protocol as the existing yaw datasets. Source: Donghyun1228/libero-support-random-waypoint-300k-20261015 at 32ed215baf01af7e82f927a62b108873aa70bd19. Four LeRobot parts are under /lerobot/medium. Use an…
Apache TsFile version of binhng/liberoobjectmaskdepthIPECCOMMUNITYformat. A Franka robot dataset for the LIBERO-Object manipulation suite, distributed in the IPEC community format with object-mask and depth observations. Each of the ten tasks asks the robot to pick up a household object and place it in the basket. The converted time series records the robot's end-effector and joint state together with the commanded action at every control step. Tasks (taskindex): 0 pick up the salad dressing and place it in the basket; 1 bbq sauce; 2 ketchup; 3 tomato sauce; 4 alphabet soup; 5 cream cheese; 6 chocolate pudding; 7 butter; 8 orange juice; 9 milk. - Time (INT64, milliseconds) - per-episode…
Nine pools across four isolated CPU notebooks. This repository owns: {"code": 6000000000, "json": 500000000, "validatedtools": 500000000}. Read latest.json, then its pinned checkpoint manifest. Counts include one EOS per document. Uint16 little-endian token banks retain document offsets and original English text in sibling Parquet indexes. Training performs its own context packing and token-aware source sampling. Index offsets are part-local. Strict natural-language checks, permissive semantic junk screening, local normalized exact deduplication. Cross-pool/prior-training deduplication and semantic/language sample review are explicitly pending. Source quality and language classifiers are…
Nine pools across four isolated CPU notebooks. This repository owns: {"educationalweb": 4000000000, "generalweb": 8000000000}. Read latest.json, then its pinned checkpoint manifest. Counts include one EOS per document. Uint16 little-endian token banks retain document offsets and original English text in sibling Parquet indexes. Training performs its own context packing and token-aware source sampling. Index offsets are part-local. Strict natural-language checks, permissive semantic junk screening, local normalized exact deduplication. Cross-pool/prior-training deduplication and semantic/language sample review are explicitly pending. Source quality and language classifiers are fallible. No…
Resin-set quartz and garnet sand packs whose CPMG was measured and whose pore space was then imaged: the relaxometry reference where the measurement and the geometry are the SAME object. Relaxation is not in the walk, so one pack answers every relaxivity. The reference it reproduces is Model Synthetic Samples for Validation of NMR Signal Simulations (Transport in Porous Media), on the same object: grade A by the rule of dmipy-sim#459. These packs disagree with the published number. 1G100: replayed 222.049 against 178.449 printed in Model Synthetic Samples for Validation of NMR Signal Simulations (Ling et al. 2022), whose released te = 100 us echo train the value is read from, the log-mean…
Held-out test wells from NoraResearchLab/Lithology-Training-Dataset. Protocol: depth-binned log summary → JSON interval prediction → sequence metrics. Figures in figures/.
Value-only chess dataset. Each row is a unique board labeled with official Stockfish 19 UCIShowWDL. Use wdl as the value target. Do not treat this as MultiPV policy data. 9,100,000 rows on this repo (append-only waves). Source id 4. Compact move vocab (1968). Shards keep a global index: wave 1 is data/shard000000–000199. Later waves continue. This is SF19's fishtest-LTC self-play WDL model (eval + remaining material). It is not FIDE/Lichess Elo and not a sigmoid of cp. Official WDL is Honor split. split=1 is a 5% holdout. Do not invent a new hash holdout. For value training, use wdl with KL / cross-entropy against a 3-class head ordered win/draw/loss. Drop or keep wdlsource==2 terminals…
Multi-vector (late-interaction) embeddings of LoTTE pooled/dev/search (lotte/pooled/dev/search), encoded with lightonai/LateOn at revision 62911e105059585d244384c7d17826e35f669c17. Source data: irdatasets lotte/pooled/dev/search (irdatasets 0.6.3), which downloads lotte.tar.gz (md5 3b2e88b1d66933627462950b4c3f5d0f). The ColBERTv2 authors also publish LoTTE on the Hub as colbertv2/lotte, whose card gives this dataset's license; the data here was loaded through irdatasets, not from that repo. Document, query and qrel ids are the source's own ids, unchanged. Every document is one variable-length set of 128-d vectors; every query is one variable-length set of 128-d vectors. Documents and…
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.