marzoukalii/llm-from-scratch-qa

LLM from Scratch QA

An English and French data set for training a small language model from scratch to answer questions. It was made for learning: the subsets follow the stages of training, so you can pretrain a model on plain text first and then fine-tune it on questions and answers.

It is the data set of the "Build an LLM from scratch" course.

Subsets and splits

Subset	Split	Rows	English	French	What it is for
pretrain	phase1_train	195,976	117,597	78,379	Phase 1: lots of text, lightly cleaned
pretrain	phase1_eval	3,999	2,400	1,599	Loss on unseen phase 1 text
pretrain	phase2_train	19,677	11,890	7,787	Phase 2: less text, strictly cleaned
pretrain	phase2_eval	1,000	600	400	Loss on unseen phase 2 text
finetune	train	93,500	89,338	4,162	Teach the model to answer questions
finetune	eval	10,995	10,580	415	Questions about unseen articles

The pretrain subset has about 172 million words: 151 million in phase 1 and 17 million in phase 2, plus 4 million in the eval splits. The order of the rows is random, and the two languages are mixed.

from datasets import load_dataset

pretrain = load_dataset("marzoukalii/llm-from-scratch-qa", "pretrain")
finetune = load_dataset("marzoukalii/llm-from-scratch-qa", "finetune")

Question types in finetune

Type	Answer	Train	Eval	Source
factual	A short span from the paragraph, such as a name or date	91,051	10,723	SQuAD, PIAF
explanatory	2 to 3 sentences on what something is or how it works	1,225	136	Claude
analytical	2 to 4 sentences reasoning from two or more facts	1,224	136	Claude

Columns

pretrain

Column	Description
id	en-wiki-… for a Wikipedia article, en-qa-… for a question article (see below)
lang	en or fr
title	Article title
text	Article text, paragraphs separated by a blank line
url	The article on Wikipedia

finetune

Column	Description
id	Unique id, such as en-squad-…, fr-piaf-… or en-qa-475-gen-1
lang	en or fr
title	Title of the article the question is about
context	The paragraph that holds the answer
question	The question
answer	The answer
type	factual, explanatory or analytical
source	rajpurkar/squad, AgentPublic/piaf or the Claude model that wrote it

An explanatory row from finetune/train:

{
  "id": "en-qa-475-gen-1",
  "lang": "en",
  "title": "War on Terror",
  "question": "What was Operation Smiles in the U.S. military mission in the Philippines?",
  "answer": "Operation Smiles was the second portion of the U.S.-assisted operation in the Philippines, run as a humanitarian program. Its goal was to provide medical care and services to the region of Basilan as part of a \"Hearts and Minds\" program. It accompanied military efforts to remove Abu Sayyaf and Jemaah Islamiyah from their stronghold on the island.",
  "type": "explanatory",
  "source": "claude-opus-5-5"
}

How it was built

The build scripts use a fixed random seed (42), so the same sources give the same splits.

1. Wikipedia articles: two random shards per language from the 20231101.en and 20231101.fr dumps of wikimedia/wikipedia. Duplicate texts and disambiguation pages are skipped, and no article is in more than one split.
2. Phase 1 cleaning is light on purpose. Only stubs under 50 words are dropped. Lists, headings and reference sections stay in, so the model sees messy text in its first phase.
3. Phase 2 cleaning is strict. An article must open with a clean paragraph, and at least 70% of its prose paragraphs must be clean: 12 or more words, a capital letter at the start, the end of a sentence at the end, and no gaps left by removed templates. Headings, lists and everything after "References", "See also", "Voir aussi" and similar sections are removed. What's left must be 300 to 5,000 words, and "List of" pages are skipped.
4. Factual questions: all of SQuAD (train and validation) for English and PIAF for French. Questions repeated word for word are dropped, and questions with several answers keep the most common one.
5. Question articles go into pretraining. SQuAD and PIAF have 681 articles between them. Each one's paragraphs become a phase2_train row with an en-qa-… or fr-qa-… id, and any Wikipedia row with the same title is removed. So the model reads the facts during pretraining before it is asked about them.
6. Explanatory and analytical questions: Claude (claude-opus-5-5) read each question article and wrote 2 explanatory and 2 analytical questions in the article's language, using only the article. Each answer came with a sentence quoted from the article. A question was kept only if its quote is in the article word for word, and that paragraph became its context. 2 of 2,724 were dropped this way, and 1 that repeated a SQuAD question.
7. Eval holds out whole articles. 10% of the question articles in each language are eval articles, and all of their questions, factual and generated, are in eval. No eval question is about an article in train.

Limitations

• Mostly English questions. SQuAD is much larger than PIAF, so 96% of the fine-tune questions are in English. The pretrain text is about 60% English and 40% French.
• Factual answers are short spans copied from the paragraph, so a model trained only on them learns to answer in a few words. The explanatory and analytical questions teach full-sentence answers, but there are far fewer of them.
• The facts are from Wikipedia. SQuAD's paragraphs come from 2016 articles and PIAF's from French Wikipedia; facts may be out of date, and the articles carry Wikipedia's coverage and biases.
• Generated answers were checked automatically, not by people. The quote check makes sure each answer is tied to text in the article, but it can't prove every sentence of the answer is correct.
• Phase 1 is noisy on purpose, as described above.

Sources and license

This data set is released under CC BY-SA 4.0. If you share it or something made from it, credit the sources below and use the same license.

Source	Used for	License
Wikipedia, via wikimedia/wikipedia	Pretrain text	CC BY-SA 4.0
SQuAD v1.1	English factual questions	CC BY-SA 4.0
PIAF	French factual questions	MIT, with paragraphs from French Wikipedia (CC BY-SA)
Claude (Anthropic)	Explanatory and analytical questions	Written for this data set

The text of each Wikipedia article is the work of its contributors; the url column links to the article and its history.

PIAF license

The French factual questions come from PIAF, released by Etalab under the MIT license. Its notice:

MIT License

Copyright (c) 2020 DINUM/Etalab

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Source citations

@inproceedings{rajpurkar-etal-2016-squad,
  title     = "{SQ}u{AD}: 100,000+ Questions for Machine Comprehension of Text",
  author    = "Rajpurkar, Pranav and Zhang, Jian and Lopyrev, Konstantin and Liang, Percy",
  booktitle = "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
  year      = "2016",
  address   = "Austin, Texas",
  publisher = "Association for Computational Linguistics",
  url       = "https://aclanthology.org/D16-1264"
}

@inproceedings{keraron-etal-2020-piaf,
  title     = "Project {PIAF}: Building a Native {F}rench Question-Answering Dataset",
  author    = "Keraron, Rachel and Lancrenon, Guillaume and Bras, Mathilde and Allary, Fr{\'e}d{\'e}ric and Moyse, Gilles and Scialom, Thomas and Soriano-Morales, Edmundo-Pavel and Staiano, Jacopo",
  booktitle = "Proceedings of the 12th Language Resources and Evaluation Conference",
  year      = "2020",
  address   = "Marseille, France",
  publisher = "European Language Resources Association"
}