fahiiim/TypoFormer-Synthetic-Dataset

TypoFormer Synthetic GEC and Autocomplete Dataset

High-Precision Multi-Domain Synthetic Dataset for Grammar Error Correction, Typo Repair, and Next-Word Generation
Created and Authored by Md. Fahim Sarker Mridul
Official Model: TypoFormer on Hugging Face | Code: TyPoTaKer on GitHub

---

Dataset Summary

This dataset contains 53,000 paired sentence instances engineered specifically for training, fine-tuning, and evaluating character-aware edge language models and non-autoregressive grammatical error correction (GEC) taggers.

The dataset is partitioned into two complementary subsets:

1. Wine Engineering & Client Dialogue Subset (28,000 rows): Focuses on software engineering, technical reviews, client communication, everyday workplace coordination, punctuation, vocabulary, and clean continuations for next-word generation.
2. Grammarly & QuillBot Grade GEC Subset (25,000 rows): Focuses on cognitive homophone confusions, subject-verb agreement, casing slips, fluency refinement, preposition collocations, and pristine negative anchors.

---

Dataset Structure

Data Splits

Configuration	Split	Instances	Format	Description
default	train	47,700	Parquet / CSV	Unified training split containing both Wine and GEC data
default	validation	5,300	Parquet / CSV	Unified evaluation split (10%)
wine	train	25,200	CSV	Engineering, client dialogue, and next-word generation training
wine	validation	2,800	CSV	Engineering and client dialogue evaluation
grammarly_quillbot	train	22,500	CSV	Core grammar, typo, and fluency refinement training
grammarly_quillbot	validation	2,500	CSV	Core grammar, typo, and fluency evaluation

---

Data Fields

Every instance contains the following 10 structured fields:

Field	Type	Description	Example
id	string	Unique record identifier	WINE_006456 or GQ_000001
input_text	string	Source sentence containing simulated typos, grammatical slips, or clean text	"We verified that their planning to deploy."
target_text	string	Target ground-truth sentence	"We verified that they're planning to deploy."
error_type	string	High-level linguistic category (grammar, spelling, syntax, fluency, punctuation, clean)	grammar
error_category	string	Fine-grained error taxonomy	homophone_confusion
error_count	integer	Count of errors present (0 for clean rows)	1
correction_explanation	string	Rationale explaining the linguistic correction	"Use 'they\'re' as a contraction for 'they are'."
domain	string	Contextual writing domain	technical_code
complexity_level	string	Complexity tier (elementary, intermediate, advanced, professional)	intermediate
subset	string	Originating subset (wine_engineering_chat or grammarly_quillbot_gec)	wine_engineering_chat

---

Linguistic Taxonomy and Error Breakdown

The dataset covers a balanced distribution of real-world typing errors, cognitive mistakes, and clean negative anchors:

Error Family	Share	Linguistic Focus
Grammar	35.8%	Homophone confusion (their/there/they're, its/it's, affect/effect), verb tenses, and countability.
Clean Anchors	23.4%	Pristine negative samples (0 errors) to prevent over-correction and ensure clean text preservation.
Spelling and Typos	13.0%	QWERTY keyboard adjacency slips, letter transpositions, omissions, and fat-finger corruptions.
Fluency	12.8%	Wordiness reduction, redundant phrasing, and natural collocation alignment.
Syntax	11.4%	Subject-verb agreement, inverted clause structures, and pronoun case.
Punctuation	3.6%	Comma splices, missing apostrophes in contractions, capitalization, and sentence terminators.

---

Domain Coverage

• Software Engineering & Systems (41.2%): Distributed systems, microservices, cloud deployments, CUDA kernels, database transactions, concurrency, API endpoints, pull requests.
• Client Handling & Business Email (24.1%): Client onboarding, contract negotiations, implementation timelines, technical reviews, stakeholder updates, SLA agreements.
• Academic & Scientific (18.5%): Empirical hypotheses, statistical regression, benchmark comparisons, literature citations, research methodologies.
• Corporate Strategy & Finance (9.2%): Quarterly targets, budget allocations, operational throughput, compliance standards.
• Everyday Workplace Chat (7.0%): Meeting handoffs, quick standup notes, reminders, conversational coordination.

---

Usage with Hugging Face Datasets

Install the library:

pip install datasets pandas pyarrow

Loading the Default Unified Split

from datasets import load_dataset

dataset = load_dataset("fahiiim/TypoFormer-Synthetic-Dataset")

print("Train rows:", len(dataset["train"]))
print("Validation rows:", len(dataset["validation"]))

# Inspect first instance
sample = dataset["train"][0]
print("Input: ", sample["input_text"])
print("Target:", sample["target_text"])
print("Category:", sample["error_category"])

Loading Specific Subsets

# 1. Load Wine Engineering & Client Chat subset
wine_data = load_dataset("fahiiim/TypoFormer-Synthetic-Dataset", "wine")

# 2. Load Grammarly & QuillBot Grade GEC subset
gec_data = load_dataset("fahiiim/TypoFormer-Synthetic-Dataset", "grammarly_quillbot")

---

Generation Methodology

All instances were programmatically synthesized and validated using custom deterministic perturbation engines located in the generation_scripts/ directory:

1. generate_wine_dataset.py:

   • Synthesizes 28,000 rows across engineering, client negotiation, and workplace dialogue.
   • Generates exact clean sentence lists (wine_lm_prompts.jsonl) used for causal next-word autocomplete training.

2. generate_grammarly_dataset.py:

   • Synthesizes 25,000 rows with precise linguistic perturbation templates.
   • Enforces a 23.4% clean preservation anchor rate so models trained on this data achieve 100% fidelity on clean sentences.

---

Associated Models and Artifacts

• Model Checkpoints: fahiiim/TypoFormer
• Standalone PyTorch Weights: gec_stage2_model.safetensors and lm_stage2_model.safetensors
• Inference Engine: Sub-10ms CPU INT8 ONNX graphs (gec_stage2_int8.onnx, lm_stage2_int8.onnx)
• GitHub Repository: fahiiim/TyPoTaKer

---

Citation and Author Attribution

@dataset{typoformer_synthetic_dataset_2026,
  author       = {Md. Fahim Sarker Mridul},
  title        = {TypoFormer Synthetic GEC and Autocomplete Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/fahiiim/TypoFormer-Synthetic-Dataset},
  howpublished = {\url{https://github.com/fahiiim/TyPoTaKer}}
}

License

This dataset is distributed under the MIT License.
Authored by Md. Fahim Sarker Mridul.