datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Math-Expanded
Massive Step-by-Step Mathematics Instruction Dataset
Dataset Description
This is a 60GB, highly knowledge-dense dataset designed to teach Large Language Models (LLMs) rigorous mathematical reasoning.
Unlike standard math datasets that only provide the final answer, this dataset emphasizes Chain-of-Thought (CoT) reasoning. Every single row contains a detailed, step-by-step breakdown of how to arrive at the solution, making it ideal for supervised fine-tuning (SFT)… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Math-Expanded.English-Mini
LLM-English-100MB — Compact & Dense English Teaching Corpus
A 100MB, extremely clean CSV designed to teach an LLM English from scratch via instruction-tuning. No noise, no HTML, no duplicates — just pure grammar, vocabulary, and syntax transformations.
Generated with a single paste-and-run Python script in Google Colab.
Why this teaches English
Instead of raw text, the dataset is instruction -> input -> output pairs that force the model to learn rules:
Grammar… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/English-Mini.CANADA_ACT_REGULATION_QA
Canadian Acts and Regulation QA
source- https://laws-lois.justice.gc.ca/eng/XML/Legis.xml
model_name="gemini-1.5-flash-latest" with 1 million context length,
First summarize the text scrapped text from xml tree of urls using gemini.
then generate QA from sumarised text.
Performance of Gemini was way way better than GPT-4.
Fitering was done based on Heuristics after rigrous analysis because llms were not always accurate.
summary_prompt_template= """
You'r legal expert… See the full description on the dataset page: https://huggingface.co/datasets/Guggu/CANADA_ACT_REGULATION_QA.Pattern-Recognition
Pattern Completion Dataset
A 30 GB synthetic dataset of numeric sequence‑completion prompts and their next values, designed to teach large language models how to recognize and extrapolate patterns.
Each row contains a prompt (the sequence with a ? indicating the missing next element) and a completion (the correct next number).
Dataset Structure
Format: CSV (no header row)
Columns:
prompt – "Find the next number in the sequence: a,b,c,... ,?"
completion – the… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Pattern-Recognition.Math-Dataset
Math Teaching Dataset - 100MB / 735k Rows
A large math dataset designed for training and fine-tuning AI models with step-by-step reasoning.
Files
math_dataset_100MB.csv - Full dataset, 100.33 MB, 735,000 rows
math_dataset_100MB.csv.gz - Compressed version, 18.95 MB
splits/math_part_1.csv to math_part_10.csv - 10 x ~10MB chunks
Dataset Stats
Total rows: 735,000
Size: 100.33 MB raw, 18.95 MB gzipped
Format: CSV, UTF-8, with header… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Math-Dataset.Math-Mini
Clean Math Reasoning Dataset
A clean dataset for training and evaluating language models on mathematical problem solving.
The dataset contains concise mathematical question-and-answer pairs designed to improve model performance on structured numerical reasoning tasks.
Dataset Structure
The dataset contains two fields:
Column
Description
prompt
A mathematical problem or question
response
The corresponding solution
Example:
prompt:
48392+92831=?… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Math-Mini.CoT-for-LLM
README — Advanced Chain‑of‑Thought Dataset Generator
Overview
This project generates a large-scale synthetic dataset of Chain‑of‑Thought (CoT) reasoning examples across multiple domains:
Math (algebra, word problems, multi‑step reasoning)
English (vocabulary explanations, nuance, tone)
Writing (multi‑paragraph reflections, structured planning)
Coding (advanced algorithms, data structures, real code snippets)
Science (physics, biology, chemistry, earth science… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/CoT-for-LLM.English
Synthetic English Language Acquisition Dataset (3GB)
A structured, 3GB synthetic CSV dataset generated to assist in pretraining or fine-tuning Language Models (LLMs) on core English syntax, vocabulary, narrative structures, and explicit grammar rules.
Dataset Structure
The dataset contains four primary columns:
Column Name
Data Type
Description
data_type
string
Categorises the entry (sentence, narrative_chunk, or grammar_instruction).
text_content… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/English.Code-Syntax-Expanded
Code-Syntax-Expanded
A massive, high-quality synthetic dataset for training LLMs to identify and correct syntax errors across 33 programming languages. Contains 5+ million unique examples (~1.1 GB) with English explanations – no artificial padding, no duplicate rows.
📊 Dataset Overview
Property
Value
Total rows
5,000,000+
File size
~1.1 GB (uncompressed CSV)
Languages
33
Unique templates
160+ error patterns
Format
CSV (4 columns)
License… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Code-Syntax-Expanded.Legal
Massive Legal & Law Instruction Tuning Dataset (5GB)
This dataset contains a massive, synthetically generated corpus of legal instruction-tuning data designed to teach Large Language Models (LLMs) how to understand, reason through, and respond to legal queries. The dataset scales progressively from fundamental legal concepts to highly advanced jurisprudential masterclasses.
Dataset Structure
The dataset is provided as a single CSV file… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Legal.Pretrain
pretrain-60GB
A 60GB, knowledge-dense pretraining corpus built as a clean, superior alternative to TinyStories.
TinyStories teaches grammar. This teaches knowledge, reasoning, and code.
File: pretrain.csv - 60GB, ~15M rows, single column text
Why vs TinyStories?
TinyStories is fiction for kids. This is textbook / encyclopedia / reasoning / code. Every row is 3500-5200 chars packed with 2-3 fused concepts, not a story.
Physics, Chem, Bio, Math, CS, History, Geo… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Pretrain.English-Extended
English Teaching Dataset
A massive, clean dataset designed to teach English to Large Language Models.
Dataset Description
Size: 20 GBRows: ~85 million examplesFormat: CSVPurpose: English language instruction fine-tuning for LLM development
This dataset provides structured, pedagogical English teaching examples optimised for training language models. Each entry follows consistent formatting to ensure clean, predictable learning patterns for AI systems.… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/English-Extended.Code-Syntax
Code Syntax Dataset (S)
A large-scale, high‑quality dataset for teaching large language models to identify and correct common syntax errors across 30+ programming languages.Contains 500,000+ unique examples (≈110 MB) with English explanations – no artificial padding.
📊 Dataset Format
The dataset is provided as a single CSV file with the following columns:
Column
Type
Description
wrong_code
string
Code snippet containing a syntax error
correct_code… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Code-Syntax.Token-Efficiency
token_efficiency_corpus
A 2.5 GB CSV corpus teaching LLMs to minimize token usage in their outputs.
Progresses from basic filler removal to expert-level nested reasoning compression.
Contents
verbose_output - The padded, wasteful version of the text
efficient_output - The compressed, token-efficient equivalent
technique - Compression strategy used
subcategory - Specific variant of the technique
difficulty - Tier 1 (easiest)… See the full description on the dataset page: https://huggingface.co/datasets/Gugu8/Token-Efficiency.Breaking_Bad_Scenes_LLM