datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
coedit
Dataset Card for CoEdIT: Text Editing via Instruction Tuning
Paper: CoEdIT: Text Editing by Task-Specific Instruction Tuning
Authors: Vipul Raheja, Dhruv Kumar, Ryan Koo, Dongyeop Kang
Project Repo: https://github.com/vipulraheja/coedit
Dataset Summary
This is the dataset that was used to train the CoEdIT text editing models. Full details of the dataset can be found in our paper.
Dataset Structure
The dataset is in JSON format.… See the full description on the dataset page: https://huggingface.co/datasets/grammarly/coedit.european-open-data-catalogue
European Open Data Catalogue
This repository publishes independently versioned metadata and licensed source snapshots:
A discovery catalogue with 15565 dataset entries from
ISTAT, Eurostat, OECD, ILO, DoveVannoINostriSoldi (DVNS) and Cruscotto Italia.
3 independently pinned availability indexes with
911,795 joint combinations across 35 datasets, built from complete
source responses within the explicitly declared scope.
Licensed Cruscotto source snapshots, stored separately from… See the full description on the dataset page: https://huggingface.co/datasets/Gramscii-IT/european-open-data-catalogue.gramvaani_preprocessed_hi_train
Dataset Card for "gramvaani_preprocessed_hi_train"
More Information needed
essay-grammar-range-qwen3.5-4b-trl-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/essay-grammar-range-qwen3.5-4b-grpo.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion
completion: the… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/essay-grammar-range-qwen3.5-4b-trl-completions.grammar-accuracy-qwen3.5-4b-trl-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/grammar-accuracy-qwen3.5-4b-grpo.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion
completion:… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-completions.unified-grammar
Unified Distributional Grammar (Greek + Latin + Hebrew)
NuBerea/unified-grammar — the j-layer construction × function × source × era matrix, mirroring the
shape of NuBerea/distributional-lexicon (lemma × sense × source × era) for grammar instead of
lexicon: every grammatical claim scoped, counted, basis-carrying (see METRIC SEMANTICS below), and traceable to corpus
instances, with traditional grammars (Smyth, Gesenius-Kautzsch-Cowley, Allen & Greenough) admitted only
as witness… See the full description on the dataset page: https://huggingface.co/datasets/NuBerea/unified-grammar.english-grammaticality
English Grammaticality Dataset
The English Grammaticality Dataset contains English sentences and their grammaticality scores as rated by native speakers.
This dataset is a combination of two other datasets listed in the references.
Composition
The dataset consists of 7350 sentences. Each sentence is represented by a row in the JSONLines format, which includes the following fields:
text: The English language sentence.
score: The average grammaticality of the sentence as… See the full description on the dataset page: https://huggingface.co/datasets/agentlans/english-grammaticality.allen-greenough-grammar
allen-greenough-grammar
Allen and Greenough's New Latin Grammar for Schools and Colleges (J.B.
Greenough, G.L. Kittredge, A.A. Howard, Benjamin L. D'Ooge, eds.; Boston:
Ginn & Company, 1903 — public domain), in the Dickinson College
Commentaries (DCC) digital re-edition
(https://dcc.dickinson.edu/grammar/latin/), edited by Meagan Ayer under
Chris Francese's direction, 2013-2016 ("A New Allen and Greenough"). This is
a PROSE-witness t0 source repo, the standard reference grammar… See the full description on the dataset page: https://huggingface.co/datasets/NuBerea/allen-greenough-grammar.smyth-grammar
smyth-grammar
Herbert Weir Smyth, A Greek Grammar for Colleges (New York: American Book
Company, 1920 — public domain). PROSE-witness t0 source repo for Greek
morphology and syntax, the sibling of allen-greenough-grammar (Latin) and
gesenius-kautzsch-grammar (Hebrew) in the distributional-grammar programme:
one row per numbered Smyth paragraph (§1–§3048, complete, plus the 213
"D"-suffixed dialect paragraphs), Greek examples in polytonic Unicode, Smyth's
own cross-references as… See the full description on the dataset page: https://huggingface.co/datasets/NuBerea/smyth-grammar.gesenius-kautzsch-grammar
gesenius-kautzsch-grammar
Gesenius' Hebrew Grammar, as edited and enlarged by E. Kautzsch, 28th
German edition (1909), translated by A. E. Cowley — Second English
Edition (Oxford: Clarendon Press, 1910; public domain), in the
Wikisource collaborative transcription
(https://en.wikisource.org/wiki/Gesenius%27_Hebrew_Grammar). This is a
PROSE-witness t0 source repo, the standard reference grammar for Biblical
Hebrew morphology and syntax cited throughout the NuBerea Hebrew corpus… See the full description on the dataset page: https://huggingface.co/datasets/NuBerea/gesenius-kautzsch-grammar.grammar-correction
grammar-correction
Dataset Summary
The grammar-correction dataset is a refined subset of the liweili/c4_200m dataset,
derived from Google's C4_200M Synthetic Dataset for Grammatical Error Correction.
It contains sentence pairs where the input is ungrammatical and the output is grammatical, making it suitable for training grammatical error correction (GEC) models.
Dataset Structure
Train set: 100 000 entries
Validation set: 25 000 entries… See the full description on the dataset page: https://huggingface.co/datasets/agentlans/grammar-correction.texts-labelled-grammaticalitymedit
Dataset Card for mEdIT: Multilingual Text Editing via Instruction Tuning
Paper: mEdIT: Multilingual Text Editing via Instruction Tuning
Authors: Vipul Raheja, Dimitris Alikaniotis, Vivek Kulkarni, Bashar Alhafni, Dhruv Kumar
Project Repo: https://github.com/vipulraheja/medit
Dataset Summary
This is the dataset that was used to train the mEdIT text editing models. Full details of the dataset can be found in our paper.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/grammarly/medit.grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-grpo-vllm-colocate-completions.french-bench-grammar-vocab-reading
Dataset Card for "french-bench-grammar-vocab-reading"
More Information needed
spivavtor
Dataset Card for Spivavtor
Paper: Spivavtor: An Instruction Tuned Ukrainian Text Editing Model
Authors: Aman Saini, Artem Chernodub, Vipul Raheja, Vivek Kulkarni
Dataset Summary
This is the dataset used to train all Spivavtor models. It contains data for 4 tasks - Grammatical Error Correction (GEC), Simplification, Coherence and Paraphrasing.
The specific details are as follows:
Task
Examples in Training data
Examples in Validation… See the full description on the dataset page: https://huggingface.co/datasets/grammarly/spivavtor.european-public-funding
European public funding
A map of where public bodies in Italy, and the European Union, publish their
calls for grants and funding to businesses (in Italian, bandi; procurement
tenders are left out on purpose): which body, on which interface, at which
listing, and what the last verification of that interface answered. It holds
no calls. Gramscii's Bandi collector reads the calls from these places into
its own catalogue; this dataset is the map it reads them by. Every file is… See the full description on the dataset page: https://huggingface.co/datasets/Gramscii-IT/european-public-funding.1000_english-grammar-datasetsemantic-repair-routing
semantic-repair-routing
The 84,819 supervised pairs that trained
SemanticRepair-270M:
a message somebody actually wrote, and the requests inside it restated
plainly, one per line.
It teaches one narrow thing. An embedding router compares a question with
the description of every capability it can reach. People do not write the
way capabilities are described — they hedge, they apologise, they ask two
things in one breath, they name what they do not want. This data pairs
the first… See the full description on the dataset page: https://huggingface.co/datasets/Gramscii-IT/semantic-repair-routing.kurdish-grammar-eval
Kurdish Grammar Minimal Pairs (BLiMP-style) — Kurmancî · Soranî
A grammar-competence benchmark for Kurdish, built on the BLiMP
idea: for each item, a correct sentence is paired with a corrupted version where one specific
grammar rule has been deliberately broken. Score a language model by checking whether it assigns higher
likelihood to the correct sentence than the corrupted one — accuracy well above 50% means the model
learned the rule, not just surface fluency.
Built by… See the full description on the dataset page: https://huggingface.co/datasets/kurdish-tech/kurdish-grammar-eval.RTSP_FTP_SMTP_DNS_PROTOCOL_GRAMMAR_DATASETkurdish-kurmanji-grammar-error-correctionThis dataset is for developing and evaluating grammatical error correction (GEC) models,
like Grammarly, for Kurdish Kurmanji. Incorrect sentences were manually collected
from YouTube comment sections of Kurdish videos and X(Twitter) and Muzaffer Cıkay added their corrections.
The source videos are documented in the source.txt file.
Usage
from datasets import load_dataset
dataset = load_dataset("muzaffercky/kurdish-kurmanji-typo-correction", split="train")
print(dataset)
detexd-benchmark
Dataset Card for DeTexD: A Benchmark Dataset for Delicate Text Detection
Dataset Summary
We define delicate text as any text that is emotionally charged or potentially triggering such that engaging with it has the potential to result in harm. This broad term covers a range of sensitive texts that vary across four major dimensions: 1) riskiness, 2) explicitness, 3) topic, and 4) target.
This dataset contains texts with fine-grained individual annotator labels from 0 to 5… See the full description on the dataset page: https://huggingface.co/datasets/grammarly/detexd-benchmark.delexicalized_n_gramsnouns-proposals
Nouns DAO Proposals Dataset
A dataset of all Nouns DAO governance proposals with pass/fail outcomes and LLM-assigned categories, designed for text classification and DAO governance research.
Dataset Description
This dataset contains all 983 proposals submitted to Nouns DAO from its inception (August 2021) through July 2026. Each proposal includes its title, full description, on-chain outcome (passed/failed), and a category label assigned by an LLM classifier.
Key… See the full description on the dataset page: https://huggingface.co/datasets/gramajo/nouns-proposals.GRAM-fine-tuning-65kThis is the dataset for Fine-tuning GRAM.
Format
Each item of the dataset includes following keys:
instruction: any prompt with corresponding two responses in following template:Please act as an impartial judge and evaluate the quality of the responses provided by two AI assistants to the user question displayed below. You should choose the assistant that follows the user's instructions and answers the user's question better.
Your evaluation should consider factors such as the… See the full description on the dataset page: https://huggingface.co/datasets/NiuTrans/GRAM-fine-tuning-65k.Reber-GrammarA Reber sequence is a grammar string made of finite states, in simple words, formed by using a confined set of characters. In the research paper that proposed the LSTM, the authors use embedded Reber grammar due to its short time lags.
Here is the chapter from my Master's thesis that briefly explains the dataset, and also includes various plots about its statistics.
Data visualiations available at about_reber.
Master's thesis: Investigating Sparsity in Recurrent Neural Networks… See the full description on the dataset page: https://huggingface.co/datasets/harshildarji/Reber-Grammar.grammar_logic_rhetoric_and_mathformal-grammar-llm-benchmark
A formal grammar benchmark to study learning and memorization of large language models
Usage
Load a specific grammar (config) and split:
from datasets import load_dataset
ds = load_dataset(
"<username>/formal-grammar-llm-benchmark",
name="pcfg_cfg3b_eq_len_skewed_prob",
)
# available splits per grammar:
# train_sequences, test_sequences, non_grammatical_sequences,
# non_grammatical_*_grammar_edit_*, non_grammatical_*_edit_distance_*… See the full description on the dataset page: https://huggingface.co/datasets/hoisoserious/formal-grammar-llm-benchmark.grammar-correctionBasically used in Correctness Chorus to train T5 model to predict grammar correction.