datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
flores_101One of the biggest challenges hindering progress in low-resource and multilingual machine translation is the
lack of good evaluation benchmarks. Current evaluation benchmarks either lack good coverage of low-resource
languages, consider only restricted domains, or are low quality because they are constructed using
semi-automatic procedures. In this work, we introduce the FLORES evaluation benchmark, consisting of 3001
sentences extracted from English Wikipedia and covering a variety of different topics and domains.
These sentences have been translated in 101 languages by professional translators through a carefully
controlled process. The resulting dataset enables better assessment of model quality on the long tail of
low-resource languages, including the evaluation of many-to-many multilingual translation systems, as all
translations are multilingually aligned. By publicly releasing such a high-quality and high-coverage dataset,
we hope to foster progress in the machine translation community and beyond.clean_mc4_itA thoroughly cleaned version of the Italian portion of the multilingual
colossal, cleaned version of Common Crawl's web crawl corpus (mC4) by AllenAI.
Based on Common Crawl dataset: "https://commoncrawl.org".
This is the processed version of Google's mC4 dataset by AllenAI, with further cleaning
detailed in the repository README file.qe4pe
Quality Estimation for Post-Editing (QE4PE)
For more details on QE4PE, see our paper and our Github repository
Gabriele Sarti • Vilém Zouhar • Grzegorz Chrupała • Ana Guerberof Arenas • Malvina Nissim • Arianna Bisazza
Word-level quality estimation (QE) detects erroneous spans in machine translations, which can direct and facilitate human post-editing. While the accuracy of word-level QE systems has been assessed extensively, their usability and downstream influence on the… See the full description on the dataset page: https://huggingface.co/datasets/gsarti/qe4pe.controlled-datagsat-vocab-sentences-tts
GSAT Vocabulary TTS Audio
Text-to-speech audio files for GSAT (General Scholastic Ability Test) English vocabulary.
Structure
audio/ - MP3 audio files organized by hash prefix (e.g., audio/ab/abcd1234....mp3)
index.jsonl - Index file mapping hashes to text and TTS engine used
Engines
Kokoro (af_heart voice) - Used for lemmas (single words/phrases)
Supertonic (M1 voice) - Used for example sentences
Audio Format
Format: MP3
Sample rate: 24kHz… See the full description on the dataset page: https://huggingface.co/datasets/TCabbage/gsat-vocab-sentences-tts.GSA-PT-Qwen2-7B-Instruct-chunk4-chunk4-data
GSA-PT-Qwen2-7B-Instruct-chunk4-chunk4-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk4-chunk4.
Each sample is tokenized and formatted with GSA gist tokens for continue pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-PT-Qwen2-7B-Instruct-chunk4-chunk4 — model trained on this dataset
GSA-PT-Llama-3.2-1B-chunk8-data
GSA-PT-Llama-3.2-1B-chunk8-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models with chunk size chunk8.
Each sample is tokenized and formatted with GSA gist tokens for continued pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Models
yuzhenm/GSA-PT-Llama-3.2-1B-chunk8 — model trained on this dataset
vigorl_datasets
ViGoRL Datasets
This repository contains the official datasets associated with the paper "Grounded Reinforcement Learning for Visual Reasoning (ViGoRL)", by Gabriel Sarch, Snigdha Saha, Naitik Khandelwal, Ayush Jain, Michael J. Tarr, Aviral Kumar, and Katerina Fragkiadaki.
Dataset Overview
These datasets are designed for training and evaluating visually grounded vision-language models (VLMs).
Datasets are organized by the visual reasoning tasks described in the ViGoRL… See the full description on the dataset page: https://huggingface.co/datasets/gsarch/vigorl_datasets.details_GSAI-ML__LLaDA-8B-Instruct_v2
Dataset Card for Evaluation run of GSAI-ML/LLaDA-8B-Instruct
Dataset automatically created during the evaluation run of model GSAI-ML/LLaDA-8B-Instruct.
The dataset is composed of 116 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An… See the full description on the dataset page: https://huggingface.co/datasets/OALL/details_GSAI-ML__LLaDA-8B-Instruct_v2.mt_genevalThe MT-GenEval benchmark evaluates gender translation accuracy on English -> {Arabic, French, German, Hindi, Italian,
Portuguese, Russian, Spanish}. The dataset contains individual sentences with annotations on the gendered target words,
and contrastive original-invertend translations with additional preceding context.GSA-PT-Qwen2-7B-Instruct-chunk16-data
GSA-PT-Qwen2-7B-Instruct-chunk16-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk16.
Each sample is tokenized and formatted with GSA gist tokens for continue pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-PT-Qwen2-7B-Instruct-chunk16 — model trained on this dataset
GSA-FT-Qwen2-7B-Instruct-chunk4-chunk4-data
GSA-FT-Qwen2-7B-Instruct-chunk4-chunk4-data
This is the supervised fine-tuning dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk4-chunk4.
Each sample is tokenized and formatted with GSA gist tokens for supervised fine-tuning.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-FT-Qwen2-7B-Instruct-chunk4-chunk4 — model trained on this dataset
iwslt2017_contextThe IWSLT 2017 Multilingual Task addresses text translation, including zero-shot translation, with a single MT system across all directions including English, German, Dutch, Italian and Romanian. As unofficial task, conventional bilingual text translation is offered between English and Arabic, French, Japanese, Chinese, German and Korean.GSA-PT-Llama-3.2-1B-chunk16-data
GSA-PT-Llama-3.2-1B-chunk16-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models with chunk size chunk16.
Each sample is tokenized and formatted with GSA gist tokens for continued pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Models
yuzhenm/GSA-PT-Llama-3.2-1B-chunk16 — model trained on this dataset
change_itThe CHANGE-IT dataset contains approximately 152,000 article-headline pairs, collected from two Italian
newspapers situated at opposite ends of the political spectrum, namely la Repubblica (left) and
Il Giornale (right), with the two newspapers equally represented. The dataset has been used in the context
of the CHANGE-IT task (https://sites.google.com/view/change-it) during the Evalita 2020 evaluation campaign
(http://www.evalita.it/2020). CHANGE-IT is a generation task for Italian – more specifically, a style transfer
task for headlines of Italian newspapers. Given a (collection of) headlines from one newspaper, namely
Il Giornale (G) or La Repubblica (R), it challenges automatic systems to change all G-headlines to headlines in
style R, and all R-headlines to headlines in style G. Although the task only concerns headline change, the dataset
comprehends both the headlines as well as their respective full articles.GSA-FT-Qwen2-7B-Instruct-chunk8-chunk4-data
GSA-FT-Qwen2-7B-Instruct-chunk8-chunk4-data
This is the supervised fine-tuning dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk8-chunk4.
Each sample is tokenized and formatted with GSA gist tokens for supervised fine-tuning.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-FT-Qwen2-7B-Instruct-chunk8-chunk4 — model trained on this dataset
GSA-PT-Qwen2-7B-Instruct-chunk32-data
GSA-PT-Qwen2-7B-Instruct-chunk32-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk32.
Each sample is tokenized and formatted with GSA gist tokens for continue pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-PT-Qwen2-7B-Instruct-chunk32 — model trained on this dataset
countqa_lite
gsarch/countqa_lite
A deterministic lite evaluation subset of Jayant-Sravan/CountQA.
Source revision: f92cc6fe46542c61e2916e3d2ae9a911e2216b1a
Source split: test
Sampling seed: 43
Output rows: 500
Schema: unchanged from the upstream dataset
CountQA is sampled at the QA-pair level. Each output row retains the original schema and contains one-element questions and answers lists, so lmms-eval's existing countqa_process_docs produces exactly 500 prompts.
Generated by… See the full description on the dataset page: https://huggingface.co/datasets/gsarch/countqa_lite.TCGA_CancermagpieThe MAGPIE corpus is a large sense-annotated corpus of potentially idiomatic expressions (PIEs), based on the British National Corpus (BNC). Potentially idiomatic expressions are like idiomatic expressions, but the term also covers literal uses of idiomatic expressions, such as 'I leave work at the end of the day.' for the idiom 'at the end of the day'. This version of the dataset reflects the filtered subset used by Dankers et al. (2022) in their investigation on how PIEs are represented by NMT models. Authors use 37k samples annotated as fully figurative or literal, for 1482 idioms that contain nouns, numerals or adjectives that are colours (which they refer to as keywords). Because idioms show syntactic and morphological variability, the focus is mostly put on nouns. PIEs and their context are separated using the original corpus’s word-level annotations.wmt_vatThe Variance-Aware Machine Translation corpus contains 70 small and discriminative test sets for machine translation (MT)
evaluation called variance-aware test sets (VAT), covering 35 translation directions from WMT16 to WMT20 competitions.
VAT is automatically created by a novel variance-aware filtering method that filters the indiscriminative test instances
of the current MT benchmark without any human labor. Experimental results show that VAT outperforms the original WMT benchmark
in terms of the correlation with human judgment across mainstream language pairs and test sets. Further analysis on the properties
of VAT reveals the challenging linguistic features (e.g., translation of low-frequency words and proper nouns) for the competitive
MT systems, providing guidance for constructing future MT test sets.ScreenSpot-Pro-Lite
ScreenSpot-Pro-Lite
A fixed 500-example representative/challenging subset of the 1,581-example
ScreenSpot-Pro benchmark.
Selection
Sampling is proportional over the cross-product of platform, application, and UI type,
so all 26 applications remain represented and the original icon/text mix is retained.
Within every stratum, 80% is deterministic seeded sampling and 20% is a hard tier.
Hardness combines failure rate and disagreement across five full-run anchor… See the full description on the dataset page: https://huggingface.co/datasets/gsarch/ScreenSpot-Pro-Lite.GSA-PT-Qwen2-7B-Instruct-chunk8-chunk4-data
GSA-PT-Qwen2-7B-Instruct-chunk8-chunk4-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk8-chunk4.
Each sample is tokenized and formatted with GSA gist tokens for continue pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-PT-Qwen2-7B-Instruct-chunk8-chunk4 — model trained on this dataset
GSA-FT-Qwen2-7B-Instruct-chunk32-data
GSA-FT-Qwen2-7B-Instruct-chunk32-data
This is the supervised fine-tuning dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk32.
Each sample is tokenized and formatted with GSA gist tokens for supervised fine-tuning.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-FT-Qwen2-7B-Instruct-chunk32 — model trained on this dataset
turkihs
Dataset Card for Tokenization Robustness
TokSuite Benchmark (Turkish Collection)
Dataset Description
This dataset is part of TokSuite, a comprehensive benchmark designed to measure how different tokenization strategies affect language model performance and robustness. This specific subset contains Turkish language multiple-choice text completion questions with various real-world perturbations that test tokenizer robustness.
Curated by: R3 Research Team… See the full description on the dataset page: https://huggingface.co/datasets/gsaltintas/turkihs.grote-logs
Dataset Card for Dataset Name
Dataset Details
Dataset Description
Curated by: [More Information Needed]
Funded by [optional]: [More Information Needed]
Shared by [optional]: [More Information Needed]
Language(s) (NLP): [More Information Needed]
License: [More Information Needed]
Dataset Sources [optional]
Repository: [More Information Needed]
Paper [optional]: [More Information Needed]
Demo [optional]: [More Information Needed]… See the full description on the dataset page: https://huggingface.co/datasets/gsarti/grote-logs.GSA-PT-Qwen2-7B-Instruct-chunk8-data
GSA-PT-Qwen2-7B-Instruct-chunk8-data
This is the continue pretraining dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk8.
Each sample is tokenized and formatted with GSA gist tokens for continue pretraining.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-PT-Qwen2-7B-Instruct-chunk8 — model trained on this dataset
ReFusion
ReFusion
Dataset Summary
This dataset is the training corpus used for ReFusion, as described in our paper. It comprises approximately 3.7 million high-quality instruction tuning samples consolidated from several state-of-the-art open-source datasets. The data covers diverse domains including mathematics, coding, and general instruction following.
Composition & Sources
The dataset is constructed from the following sources:
MAmmoTH
OpenMathInstruct-2 (1M… See the full description on the dataset page: https://huggingface.co/datasets/GSAI-ML/ReFusion.temporal_expressions
Dataset Card for Tokenization Robustness
A comprehensive evaluation dataset for testing robustness of different tokenization strategies.
Dataset Details
Dataset Description
This dataset evaluates how robust language models are to different tokenization strategies and edge cases. It includes questions with multiple choice answers designed to test various aspects of tokenization handling.
Curated by: R3
Funded by [optional]: [More Information Needed]
Shared… See the full description on the dataset page: https://huggingface.co/datasets/gsaltintas/temporal_expressions.GSA-FT-Qwen2-7B-Instruct-chunk16-data
GSA-FT-Qwen2-7B-Instruct-chunk16-data
This is the supervised fine-tuning dataset used for training GSA (Gist Sparse Attention) models based on Qwen2-7B-Instruct with chunk size chunk16.
Each sample is tokenized and formatted with GSA gist tokens for supervised fine-tuning.
Paper
GSA: Gist Sparse Attention via Learnable Compression and Selective Unfolding
Related Model
yuzhenm/GSA-FT-Qwen2-7B-Instruct-chunk16 — model trained on this dataset