datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
wikipedia-en-julian
Wikipedia English - JULIAN Training Dataset
This dataset contains cleaned English Wikipedia articles used to train the JULIAN-100M language model.
Dataset Structure
Data Fields
title (string): Article title
text (string): Full article text (cleaned and formatted)
url (string): Original Wikipedia URL
language (string): Language code ("en")
Data Example
{
"title": "Artificial Intelligence",
"text": "Artificial intelligence (AI) is intelligence… See the full description on the dataset page: https://huggingface.co/datasets/JulianKrgd/wikipedia-en-julian.crisp-165b
Crisp-165B 🧼
A crisp, aggressively-cleaned 165-billion-token English pretraining corpus — pre-tokenized (GPT-2 BPE), document-aware, near-deduplicated, benchmark-decontaminated, and nested-sampleable: take the first N tokens for any budget and you get a faithful mini-version of the whole thing.
Built to train language models from ~0.3B to ~3B parameters (1× to 100× Chinchilla) without ever touching a data pipeline. Just memmap and train.
Crisp-165B is the successor to… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/crisp-165b.finmix-v1
FinMix v1
A licence-clean financial instruction dataset for the Adaption AutoScientist
challenge (Finance category). 166,564 examples merged from 12 upstream
sources, deduplicated, and split so that no source document straddles the
train/test boundary.
Every row carries source, license and source_url, so the provenance of any
example can be traced without re-running the build.
Why this exists
Most public finance instruction sets are aggregates whose declared… See the full description on the dataset page: https://huggingface.co/datasets/julian8897/finmix-v1.SynthUI-Code-2k-v1Synth UI 🎹
https://www.synthui.design
Dataset details
This dataset aims to provide a diverse collection of NextJS code snippets, along with their corresponding instructions, to facilitate the training of language models for NextJS-related tasks. It is designed to cover a wide range of NextJS functionalities, including UI components, routing, state management, and more.
This dataset consists of:
Note: The dataset is seperated into two main parts:
raw Contains only the… See the full description on the dataset page: https://huggingface.co/datasets/JulianAT/SynthUI-Code-2k-v1.wikipedia-fr-julian
Wikipedia French - JULIAN Training Dataset
Ce dataset contient les articles Wikipedia français nettoyés utilisés pour entraîner le modèle de langage JULIAN-100M.
Description du Dataset
Langue: Français
Source: Dumps Wikipedia (dernière version disponible)
Taille: 950 millions de tokens (7.3GB JSONL, ~1.5-2GB Parquet)
Format: Articles nettoyés avec titre, texte et URL
Licence: Creative Commons Attribution-ShareAlike 3.0
Structure du Dataset
Champs de… See the full description on the dataset page: https://huggingface.co/datasets/JulianKrgd/wikipedia-fr-julian.finmix-autoscientist-10k
FinMix AutoScientist 10k
A deterministic, upload-ready 10,000-row subset of
FinMix v1, created for
fast finance adaptation runs in the Adaption AutoScientist challenge.
Use with Adaption Adaptive Data
Import this Hugging Face dataset and map:
Prompt: prompt
Context: context
Completion: completion
Leave task_type, source, and group_key unmapped. They are retained for
provenance and auditing.
Fields
Field
Description
prompt
Financial… See the full description on the dataset page: https://huggingface.co/datasets/julian8897/finmix-autoscientist-10k.smoltalk-gpt2-sft
SmolTalk-GPT2-SFT
A fully tokenized version of the entire HuggingFaceTB/smoltalk
(all config) dataset — converted from raw multi-turn conversations into ready-to-train
(token_id, loss_mask) pairs using the GPT-2 BPE tokenizer (vocab=50257). Drop
it straight into a PyTorch trainer for supervised fine-tuning (SFT) of any
GPT-2-vocab language model — no template parsing, no role tagging, no chat-template
gymnastics required.
Overview
Source: HuggingFaceTB/smoltalk (all… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/smoltalk-gpt2-sft.mixed-pretrain-100b-gpt2
Mixed-Pretrain-100B-GPT2
A 100-billion-token pretrain mix tokenized with the GPT-2 BPE tokenizer
(vocab=50257) and packed into flat np.memmap-ready uint16 binary files.
Designed as a universal pretrain corpus for GPT-2-vocab models in the 500M
to 5B parameter range (1× to 200× Chinchilla optimal).
Mix Composition (actual, after topups)
Inspired by SmolLM2 / Llama-3 / DCLM recipes. Carefully balanced for:
educational density, factual breadth, scientific reasoning, math… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/mixed-pretrain-100b-gpt2.SynthUI-Code-Instruct-2k-v1Synth UI 🎹
https://www.synthui.design
Dataset details
This dataset aims to provide a diverse collection of NextJS code snippets, along with their corresponding instructions, to facilitate the training of language models for NextJS-related tasks. It is designed to cover a wide range of NextJS functionalities, including UI components, routing, state management, and more.
This dataset consists of:
Note: The dataset is seperated into two main parts:
raw Contains only the… See the full description on the dataset page: https://huggingface.co/datasets/JulianAT/SynthUI-Code-Instruct-2k-v1.guia-de-colaboracao-do-ecossistema
Guia de Colaboração do Ecossistema de Inteligência Artificial — AI Brasil
Documento fundador da colaboração no ecossistema ai.eco.br · somos.aibrasil.ai · sou.aibrasil.ai
Este repositório publica o conteúdo do Guia de Colaboração da comunidade AI Brasil: a doutrina, a jornada de participação, o vocabulário de papéis, as regras do jogo, a camada prática da plataforma e o plano editorial da revista impressa de 48 páginas. É o material-base para quem quer entender como se colabora… See the full description on the dataset page: https://huggingface.co/datasets/Julianokimura/guia-de-colaboracao-do-ecossistema.mixed-pretrain-10b-gpt2
Mixed Pretraining 10B (GPT-2 BPE)
A 10-billion-token pretraining dataset, GPT-2 BPE tokenized, assembled as a
diverse mix of web text, books, Wikipedia, code, academic papers, Q&A and
instruction-formatted conversations.
Built to train a ~500M parameter from-scratch GPT-2-style transformer (see
juliannunezb/transformer-lm-500m).
Mix
Source
Mix %
Tokens
Notes
fineweb
40.4%
4,039,999,700
reused from kjj0/fineweb10B-gpt2
fineweb_edu
15.2%
1,514,999,900
reused… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/mixed-pretrain-10b-gpt2.crisp-128b
Crisp-128B 🧼
A crisp, aggressively-cleaned 128-billion-token English pretraining corpus — pre-tokenized (GPT-2 BPE), document-aware, and nested-sampleable: take the first N tokens for any budget and you get a faithful mini-version of the whole thing.
Built for training language models from ~0.3B to ~3B parameters (1× to 50× Chinchilla) without touching a data pipeline. Just memmap and train.
Why this exists
Most open pretraining corpora make you choose: download… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/crisp-128b.personal-codex-model
Personal Codex Model Training Corpus
Overview
Personal Codex Model Training Corpus is a provenance-aware, repository-level dataset for causal
language modeling, code completion, continued pretraining, and coding assistant adaptation. It is
built from source files present in local Git repository checkouts at a defined collection point.
The dataset prioritizes broad, authentic software-engineering coverage while retaining enough
metadata to audit every emitted… See the full description on the dataset page: https://huggingface.co/datasets/JulianAT/personal-codex-model.lsmp-rural-cad
LSMP Rural CAD Dataset
This dataset contains labeled rural residence floor plans for training CAD generation models.
Dataset Structure
train.jsonl: Training data in JSON Lines format (90% of data)
eval.jsonl: Evaluation data in JSON Lines format (10% of data)
Data Format
Each sample contains:
instruction: Fixed instruction for floor plan generation
input: Plot size, room requirements, style preference, and rural residence features
output: SVG parameters for… See the full description on the dataset page: https://huggingface.co/datasets/julialovenary/lsmp-rural-cad.RLVR-Math-16k
RLVR-Math-16k
A curated math reasoning dataset for RLVR (Reinforcement Learning with Verifiable Rewards) training.
Dataset Summary
Split
Samples
train
16,384
test
842
Total
17,226
Source Datasets
train
Source
Samples
hiyouga/math12k
10,476
nlile/NuminaMath-1.5-RL-Verifiable/amc_aime
3,075
nlile/NuminaMath-1.5-RL-Verifiable/olympiads
2,833
test
Source
Samples
hiyouga/math12k
500… See the full description on the dataset page: https://huggingface.co/datasets/Julian2002/RLVR-Math-16k.Julia-Proof-Pile-2Julia-Proof-Pile-2
This dataset is part of Proof-Pile-2 dataset. This dataset is consisting of mathematical code, including numerical computing, computer algebra, and formal mathematics.
This entire dataset is in Julia language. It is slightly more than 0.5 Billion tokens. I have removed Meta data from this dataset hence you can directly use it for training purpose.
This dataset is in Jsonl format.
julia-manuals
Julia Programming Language Documentation
This dataset contains the Julia programming language documentation,
chunked using semantic parsing for pretraining language models.
Updated: 2025-09-08
Loading
from datasets import load_dataset
ds = load_dataset("json", data_files={"train": "train.jsonl"}, split="train")
Statistics
Format: JSONL with single text field per line
Chunking: Semantic structure-aware chunking
Content: Official Julia documentation and… See the full description on the dataset page: https://huggingface.co/datasets/jusjinuk/julia-manuals.midtrain-mix-5b-gpt2
Midtrain-Mix-5B-GPT2
A 5-billion-token high-quality mid-training mix, pre-tokenized with the
GPT-2 BPE tokenizer (vocab=50257) and packed into flat np.memmap-ready
uint16 binary files. Drop it straight into any GPT-2-vocab pretrain trainer
for continued pretraining / mid-training / annealing of a model that's
already seen a generic web corpus.
What is mid-training and why does it matter?
After a standard pretrain pass on web data (FineWeb, RedPajama, DCLM, etc.)… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/midtrain-mix-5b-gpt2.ultrafeedback-gpt2-dpo
UltraFeedback-GPT2-DPO
A fully tokenized version of HuggingFaceH4/ultrafeedback_binarized
converted into ready-to-train (prompt_ids, chosen_response_ids, rejected_response_ids) triples
using the GPT-2 BPE tokenizer (vocab=50257). Drop it straight into a PyTorch DPO trainer
for Direct Preference Optimization of any GPT-2-vocab language model.
Overview
Source: HuggingFaceH4/ultrafeedback_binarized — 61K preference pairs binarized from UltraFeedback (Cui et al., 2023)… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/ultrafeedback-gpt2-dpo.crisp-v8
crisp-v8
An 8B-token, pre-tokenized English pretraining corpus with a single-digit tokenizer, in two nested-sampled phases (stable + anneal). Built to pretrain crisp3b-v8, a 2.67B model, on one GPU.
The shards are gated (see Licensing and access); the tokenizer and this card are open.
What is in it
Tokens
8.0B: final_stable/ 5.60B (6 shards) + final_anneal/ 2.40B (3 shards) + val.bin 85.5M held-out
Format
uint16 token ids, documents separated by `<… See the full description on the dataset page: https://huggingface.co/datasets/juliannunezb/crisp-v8.tessera-calibration-commons
Tessera Calibration Commons
This dataset contains architecture-fingerprinted aggregate calibration
statistics produced by the Tessera tooling in llama.cpp. It excludes prompts,
completions, request logs, raw activations, model weights, and credentials.
Each contribution is independently attributable and stored as immutable
Parquet sufficient statistics so clients can merge compatible observations.
julia-focal-method
📚 julia-focal-method
This dataset is designed as a validation benchmark for evaluating unit test generation capabilities of models for Julia code, particularly in low-resource settings.
Purpose: Evaluate how well a model can generate test code, given a focal function.
Data: Each example provides:
module_name: the Julia module where the function is defined.
function_name: the name of the focal function.
focal_code: the raw source code of the function (used for context).
repo: the… See the full description on the dataset page: https://huggingface.co/datasets/dongg18/julia-focal-method.