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 2026-10-19 · How the library is built
3,247 datasets, sorted by most downloaded.
Prepared for private transfer between the user's evaluation and paper machines. This repository contains existing evaluation data, not new experiments. The two final Qwen pools are: - shared-tb21-native-text-only-xhigh-20260918: 64/89 = 71.9101%. - shared-tb21-native-residual-xhigh-20260918: 68/89 = 76.4045%. Both live pool directories are under: /inspire/l20d/project/sais-bio/public/hanchen/rm-stack/runs/tb21/jobs/. These are Qwen3.8-27B runs, with native text compaction and step2320 residual weights in the residual arm. They are not the newer GLM pools. All 178 live committed result hashes still match the sealed 2026-09-20 export. case-study.tar.gz is approximately 123 MB. It contains all…
544 teleoperated Franka episodes: 3 tactile sensors × 4 tasks, with wrist + third-person RGB, robot state, actions, and raw tactile readings. Train and val share a folder; the original split (482 / 62) is in valsplit.json. Raw tactile dimension D, per finger
Parquet conversion of the Telegram user dataset. - parquet/users/ — 146 parquet parts, ~1.2 GB total, 72,601,073 rows
This is a random dump of data, do not train with it. You have been warned.
Titan Library Arabic Arabic editions of the public-domain books on Titan-Library, produced by Titan Connect. This is a machine translation, reviewed by a second machine. Each passage was translated by one model and then compared with the English by a model from a different provider, which corrected what it found. Before publication, a program checked every paragraph for missing or added text, for leftover Latin letters and for lost numbers. Structural fidelity is enforced by
UniRec-0.1B is a unified recognition model with only 0.1B parameters, designed for high-accuracy and efficient recognition of plain text (words, lines, paragraphs), mathematical formulas (single-line, multi-line), and mixed content in both Chinese and English. It addresses structural variability and semantic entanglement by using a hierarchical supervision training strategy and a semantic-decoupled tokenizer. Despite its small size, it achieves performance comparable to or better than much larger vision-language models. - PyTorch version >= 1.13.0 - Python version >= 3.7 It is recommended to organize your working directory as follows: Download the UniRec40M dataset from Hugging Face Run the…
Uzbek audiobook recordings with machine-generated, structurally filtered transcripts. - audio/ — the raw full-length recordings, unmodified. - manifest/gold.jsonl — segments that pass every structural check. Training set. - manifest/review.jsonl — segments held back with the reason recorded. Kept, not discarded. - manifest/reject.jsonl — segments failing a hard check. - manifest/SUMMARY.json — counts, hours, and the gate thresholds actually used. audiohf is a path inside this repo, so a segment can be cut directly from audio/. text is the transcript as decoded; textnorm is case-folded and punctuation-stripped with the okina (U+02BB) preserved as a letter and apostrophe variants folded onto…
Release: Vesper Differential Photometry Handoff; Cutoff: 2031-04-11T04:00:00Z; Earliest capture: 2031-04-10T01:15:00Z; Night partitions: training, validation; Retained exposures: 4; Total net rate (milli-ADU/ms): 2858.
Dataset · Visual document retrieval
Vidore3ComputerScienceBGEm3Reranking.v2
This task is used to evaluate the reranking performance for visual document retrieval. The candidates for reranking are the top-50 pages retrieved by the BAAI/bge-m3 model. This dataset, Computer Science, is a corpus of textbooks from the openstacks website, intended for long-document understanding tasks. Original queries were created in english, then translated to french, german, italian, portuguese and spanish.This version add the OCR'ed markdown to allow for comparison across image-text, image-only and text-only models. You can evaluate an embedding model on this dataset using the following code: To learn more about how to run models on mteb task check out the GitHub repository. If you…
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.