The images are resized and then encoded with the DC-AE f32 autoencoder. The resizing is done with a bucketmanager with base resolution 512x512, minimum side length 256, maximum side length 1024, all sides are divisible by 32 ofcourse as they needed to be encoded by the DCAEf32 encoder. The captions are generated with moondream2, encoded with siglip and bert. (Bert embeddings variance is very high, so use a norm layer). The text embeddings are padded to 64 tokens, but i have provided the
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.
2,760Models
859Datasets
254Papers
1,692Publishers
5,040Sourced relationships
Updated 2026-09-18 · How the library is built
859 datasets, sorted by most downloaded.
Since LiveCodeBench is a continuously updated benchmark, we provide different versions of the dataset. Particularly, we provide the following versions of the dataset: - releasev1: The initial release of the dataset with problems released between May 2023 and Mar 2024 containing 400 problems. - releasev2: The updated release of the dataset with problems released between May 2023 and May 2024 containing 511 problems. - releasev3: The updated release of the dataset with problems released between May 2023 and Jul 2024 containing 612 problems. - releasev4: The updated release of the dataset with problems released between May 2023 and Sep 2024 containing 713 problems. - releasev5: The updated…
1,000 examples from https://huggingface.co/datasets/llamafactory/alpacagpt4en - 1,000 examples from https://huggingface.co/datasets/llamafactory/alpacagpt4zh - 300 examples from https://huggingface.co/datasets/llamafactory/glaivetoolcallen - 300 examples from https://huggingface.co/datasets/llamafactory/glaivetoolcallzh - 91 examples for identity learning - 300 examples from https://huggingface.co/datasets/cognitivecomputations/SystemChat-2.0 - 6 examples for multimodal supervised fine-tuning - 300(en)+300(zh) examples from https://huggingface.co/datasets/hiyouga/DPO-En-Zh-20k - 300 examples from https://huggingface.co/datasets/argilla/kto-mix-15k - 300 examples from…
3GPP specification mirror Part of the Open-Telco Telecom Standards Corpus.Sibling mirrors: mirrors 3GPP's published specifications. Every source is kept twice: original/ is the document as 3GPP released it, and marked/ is that same document converted to Markdown for search and retrieval, one raw.md per document with any figures extracted beside it. The two trees share identical paths, so a file in original/
EF5/CREST hourly forcing for CONUS, 2016-present, packed as one tar per variable/year. Each.tar expands to individual.pqf (Apache Arrow parquet) grids readable by the EF5 v4.5 native parquet reader. Used by the Space vincewin/CRESTAI. Download + extract one year, e.g.
configname: en datafiles: - configname: es datafiles
This is a smaller version of the datacomp1b dataset. Filtering was done by taking all rows that had self similarity (inner product) above 0.32. This resulted in 213009083 (213 million) rows. The results of the datacomp paper suggest that filtering by CLIP score is better than random sampling. Included in this repo are search indices created using autofaiss, over the text and image embeddings. There are two ways to access metadata, either in.parquet files in the./metadata directory, or the./index/metadata.hdf5 hdf5 file. I would suggest using embedding-reader to load the text and image embeddings.
This is a heavily filtered version of the UltraChat dataset and was used to train Zephyr-7B-β, a state of the art 7b chat model. The original datasets consists of 1.4M dialogues generated by ChatGPT and spanning a wide range of topics. To create UltraChat 200k, we applied the following logic: - Selection of a subset of data for faster supervised fine tuning. - Truecasing of the dataset, as we observed around 5% of the data contained grammatical errors like "Hello. how are you?" instead of "Hello. How are you?" - Removal of dialogues where the assistant replies with phrases like "I do not have emotions" or "I don't have opinions", even for fact-based prompts that don't involve either. The…
This is a standardized, high-efficiency version of the JUNJIE99/MegaPairs dataset. Why use this version? The original dataset is distributed as a massive Tar archive containing millions of images, accompanied by a separate JSONL annotation file. Each row in the dataset represents a Universal Retrieval Pair (Query -> Target). You can load this dataset directly using the datasets library. This is the easiest way. The library handles the Arrow files automatically. from datasets import loaddataset dataset = loaddataset("86Cao/MegaPairs-Standard", split="train") print(f"Total samples: {len(dataset)}") sample = dataset[0] print(f"Text: {sample['querytexts'][0]}") sample['queryimage'].show() #…
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.