Datasets:

IFM
/

Jais-3-Data

license:
  - odc-by
  - cc-by-4.0
language:
  - ar
pretty_name: Jais-3 Data
task_categories:
  - text-generation
tags:
  - jais-3
  - training-data
  - parquet
  - arabic
configs:
  - config_name: arabic-txt360-cc
    default: true
    data_files:
      - split: train
        path: cpt/arabic-txt360-cc/*.parquet
  - config_name: arabic-txt360-style-rewrite
    data_files:
      - split: train
        path: cpt/arabic-txt360-style-rewrite/*.parquet
  - config_name: arabic-cc-qa
    data_files:
      - split: train
        path: cpt/arabic-cc-qa/*.parquet
  - config_name: arabic-txt360-3efforts-sft
    data_files:
      - split: train
        path: sft/arabic-txt360-3efforts-sft/*.parquet

Jais-3 Data

Dataset Description

Jais-3 Data is a 23.8 billion-token sample of the data used to train the Jais-3 family of Arabic language models, released by the Institute of Foundation Models (IFM) at MBZUAI. It covers about 10% of the continued-pretraining (CPT) and supervised-fine-tuning (SFT) tokens of Jais-3 and is part of the Jais 3 collection. It is intended for language-model training and research; the four subsets can be streamed independently or combined with user-defined weights.

Dataset Subsets

Subset	Stage	Content	Records	Tokens	License
`arabic-txt360-cc`	CPT	Arabic web text from 83 Common Crawl dumps (CC-MAIN-2013-48 to CC-MAIN-2024-30)	39,191,869	15.07 B	ODC-BY
`arabic-txt360-style-rewrite`	CPT	The same web text rewritten in formal, simple and funny styles	12,379,664	4.42 B	ODC-BY
`arabic-cc-qa`	CPT	Web passages followed by generated question-answer pairs	2,966,538	1.16 B	ODC-BY
`arabic-txt360-3efforts-sft`	SFT	Arabic conversations with and without reasoning traces	2,811,750	3.20 B	CC BY 4.0

arabic-txt360 marks data built with the TxT360 recipe applied to Arabic web text; arabic-txt360-3efforts-sft is the Arabic counterpart of LLM360's TxT360-3efforts. None of the four is published separately. Each subset is a single train split of Parquet shards under cpt/ or sft/; MANIFEST_parquet.tsv lists the sha256, record count, size and token count of every shard. About 2.0% of arabic-cc-qa passages also appear in arabic-txt360-cc.

In arabic-txt360-cc, GlotLID labels 89.5% of documents Modern Standard Arabic, 6.7% Moroccan, 2.5% Egyptian and 1.0% Najdi Arabic; 7.7% are flagged code-switched. The rewrite subset is 49.0% funny, 42.6% formal and 8.4% simple. In the SFT subset 33.0% of conversations carry a reasoning trace, written mostly in English, across six categories: STEM, code, chat, math, teacher-as-assistant and math dialogue.

Data Fields

All subsets carry text or messages and token_count, the number of Jais-3 tokenizer tokens.

arabic-txt360-cc: meta with url, timestamp, cc-path, lang and lang_score (fastText), glotid and glotid_prob (GlotLID), is_code_switched.
arabic-txt360-style-rewrite: source_style (formal, simple or funny).
arabic-cc-qa: meta with url, timestamp, cc-path, detected_language, num_qa_pairs (0 for the 5.0% of passages whose pairs were all rejected), qa_token_ratio, skip_reason.
arabic-txt360-3efforts-sft: messages (turns with role, content, think, think_faster), category, conv_index, n_turns. Every assistant turn has exactly one of think (reasoning trace) or think_faster (empty string). About 9.9% of conversations start with a Jais-3 identity system prompt.

from datasets import load_dataset
dataset = load_dataset("IFM/Jais-3-Data", "arabic-txt360-cc", split="train", streaming=True)
print(next(iter(dataset)))

Data Provenance and Processing

This dataset has been collected, cleaned, and processed entirely by the IFM team.

The CPT subsets come from Common Crawl with Arabic language identification, quality filtering and deduplication; the rewrite and question-answer subsets are derived from that web text. The SFT subset is built from Arabic translations of prompts in LLM360's TxT360-3efforts, with model-generated responses and reasoning traces. E-mail addresses and IPv4 addresses in all four subsets are replaced by [EMAIL] and [IP]. The web-derived subsets cover Common Crawl dumps up to CC-MAIN-2024-30 (July 2024).

Limitations and Responsible Use

Large-scale training data can contain factual errors, duplicates, sensitive topics and unsafe content; users are responsible for evaluation and filtering appropriate to their application. To request the removal of content, open a discussion in this repository.

License and Terms of Use

The three CPT subsets are licensed under the Open Data Commons Attribution License (ODC-BY) v1.0; the underlying pages are subject to the Common Crawl Terms of Use. The SFT subset is licensed under CC BY 4.0 and is derived from TxT360-3efforts by LLM360, released under the same license.

Citation

@misc{ifm2026jais3introducingarabiccentric,
      title={Jais 3: Introducing Arabic-centric Agentic AI},
      author={Team IFM and Haonan Li and Hadi Abdine and Nikhil Ranjan and Mohamed Anwar and Ahmad Chamma and Amr Mohamed and Ahmed Frikha and Dani Bouch and Maaz Maqsood and George Ibrahim and Mostafa Awad and Nurkhan Laiyk and Hasan Iqbal and Abhishek Maiti and Bilal Bouardi and Shivam Chauhan and Ahmer Tabassum and Owais Aijaz and Omar Arbi and Neha Sengupta and Aaryamonvikram Singh and Abdelhakim Freihat and Sarath Chandran and Rahul Pal and Mohammed Bellam and Sajid Siddiki and Sarah AlBarri and Debopriyo Banerjee and Ali El Filali and Samta Kamboj and Rifo Genadi and Ahmed Attia and Go Inoue and Shuonan Dong and Amine Khelif Khelif and Evan Dufraisse and Florian Valade and Tapas Sen and Guowei He and Sunil Sahu and Onkar Pandit and Kareem Magdy and Sheikh Khalid bin Khalifa Al Nahyan and Samujjwal Ghosh and Zainul Abedien and Ahmed Quraishi and Larry Murray and Avraham Sheinin and Natalia Vassilieva and Preslav Nakov and Guokan Shang and Xuguang Hector Ren and Zhengzhong Liu and Eric Xing},
      year={2026},
      eprint={2611.10224},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2611.10224},
}