JibayAi/jibay_jcf_base

Jibay Persian Chat Dataset

---

📖 Overview

Jibay Persian Chat Dataset is a large-scale, conversational instruction dataset primarily written in Persian (Farsi), with a limited amount of English content mixed in. The dataset was curated and formatted specifically to be compatible with the Jibay Chat Format (JCF) — a structured, role-based prompt format designed by JibayAI for training, fine-tuning, and evaluating large language models.

Each entry in this dataset represents a complete conversation turn consisting of a user prompt and an assistant response, wrapped inside a consistent, deterministic system prompt. The result is a clean, parseable, and model-ready corpus suitable for supervised fine-tuning (SFT), instruction tuning, evaluation benchmarking, and format-conversion pipelines.

This dataset is released under the MIT License — you may use, modify, and redistribute it freely, provided that the name "Jibay" is preserved in derivative works, documentation, and redistributions.

---

🧩 Format: Jibay Chat Format (JCF)

All entries in this dataset follow the Jibay Chat Format, which is an open, human-readable, tag-based conversation format. Unlike JSONL or Parquet, JCF uses explicit XML-like role blocks that make conversations trivially parsable with simple string operations and equally readable by humans.

Each record follows this structure:

<[role_system]>
You are Jibay, developed by JibayAI. You are a helpful and safe assistant and have a natural and human-like tone.
<[role_system_end]>

<[role_user]>
{prompt}
<[role_user_end]>

<[role_assistant]>
{response}
<[role_assistant_end]>

Why JCF?

• Human-readable — no need for pretty-printers or JSON viewers.
• Deterministic parsing — role boundaries are unambiguous string tokens.
• Model-agnostic — can be converted to ChatML, Alpaca, ShareGPT, OpenAI Messages, or any other format with a simple transformer.
• Stream-friendly — records can be read line-by-line or block-by-block without loading the whole file into memory.
• Extensible — optional blocks can be added without breaking existing parsers.

For the official specification, see the Jibay Chat Format repository.

---

🌐 Language Distribution

Language	Approximate Share
Persian (Farsi)	~95–98%
English	~2–5%

The dataset is intentionally Persian-first. English entries exist only as a small minority — mostly for technical, scientific, or translation-related content. This makes the dataset particularly valuable for training or evaluating Persian-language LLMs, which are comparatively underrepresented in open datasets.

---

📚 Topics Covered

The dataset spans a broad range of domains, though the count of samples per topic is limited and not uniformly distributed. The main categories include:

• 💻 Programming & Code — PHP, Python, JavaScript, C++, HTML/CSS, SQL, shell scripting, debugging, algorithms, data structures, and general software engineering questions.
• 💬 Conversation & Chat — small talk, casual dialogue, greetings, opinions, and general chit-chat in Persian.
• ➗ Mathematics — arithmetic, algebra, geometry, calculus, statistics, probability, and basic problem-solving.
• 📜 History — Persian history, world history, dynasties, wars, key figures, and historical events.
• 🕌 Religion — Islamic studies, Quranic concepts, fiqh, hadith, and general religious discussions.
• 🗺️ Geography — countries, capitals, rivers, mountains, climate, and regional facts.
• 🏛️ Politics — political systems, ideologies, international relations, and current affairs (general, non-partisan).
• 🧠 General Knowledge — trivia, definitions, "who/what/when/where" questions.
• 🍲 Food & Cooking — Persian cuisine, recipes, ingredients, and culinary tips.
• 🌍 Translation — Persian ↔ English and Persian ↔ other language translation tasks.
• 📖 Literature — Persian poetry (Hafez, Saadi, Rumi, Ferdowsi), prose, and literary analysis.
• 🔬 Science — physics, chemistry, biology, and general science questions.
• 🎨 Miscellaneous — anything that doesn't fit neatly into the above categories.

⚠️ Note: Topic coverage is limited — this dataset is not intended to be a comprehensive benchmark across all domains. It is a curated collection and should be treated as a starting point rather than a definitive corpus.

---

⚠️ Data Quality & Cleaning Notice

Some samples and sections of this dataset require further cleaning. While the corpus has already been normalized to UTF-8 and formatted into JCF, users should be aware of the following:

• A small number of entries may contain duplicate or near-duplicate prompts.
• Some responses may contain formatting artifacts, truncated text, or incomplete sentences.
• A minority of samples contain off-topic content or low-quality responses that do not meet strict instruction-tuning standards.
• Certain entries reference specific individuals that have been removed for privacy or quality reasons.
• Whitespace, newline, and punctuation inconsistencies may exist in a subset of records.

If you plan to use this dataset for training, it is strongly recommended that you run your own cleaning, deduplication, and quality-filtering pipeline before fine-tuning.

---

🤖 Source Models

A portion of the conversations in this dataset were generated, assisted, or inspired by outputs from the following large language models:

• GPT-5 Codex
• Claude Opus 4.6
• Jibay 4.1

These models contributed to various conversations across programming, reasoning, and general chat categories. The dataset does not claim exclusive authorship of any model's outputs and is distributed as a mixed-source corpus for research and educational purposes.

---

🔧 System Prompt

Every record in this dataset uses the following fixed system prompt, as defined by the Jibay Chat Format:

You are Jibay, developed by JibayAI. You are a helpful and safe assistant and have a natural and human-like tone.

This system prompt is applied uniformly across all samples to ensure consistent behavior and reproducible model conditioning.

---

🚀 Usage Examples

Loading with Python

import re

def parse_jcf(text):
    """Parse a JCF-formatted string into a list of conversation dicts."""
    pattern = re.compile(
        r"<\[role_system\]>\s*(.*?)\s*<\[role_system_end\]>\s*"
        r"<\[role_user\]>\s*(.*?)\s*<\[role_user_end\]>\s*"
        r"<\[role_assistant\]>\s*(.*?)\s*<\[role_assistant_end\]>",
        re.DOTALL,
    )
    conversations = []
    for match in pattern.finditer(text):
        conversations.append({
            "system":   match.group(1),
            "prompt":   match.group(2),
            "response": match.group(3),
        })
    return conversations

with open("dataset_jcf.txt", "r", encoding="utf-8") as f:
    data = f.read()

conversations = parse_jcf(data)
print(f"Loaded {len(conversations)} conversations")
print(conversations[0])

Counting Entries with PHP

<?php
$file = __DIR__ . '/dataset_jcf.txt';
$in = fopen($file, 'rb');
$count = 0;
$buffer = '';

while (!feof($in)) {
    $buffer .= fread($in, 65536);
    $safe = max(0, strlen($buffer) - 32);
    if ($safe > 0) {
        $count += substr_count(substr($buffer, 0, $safe), '<[role_user]>');
        $buffer = substr($buffer, $safe);
    }
}
if ($buffer !== '') $count += substr_count($buffer, '<[role_user]>');
fclose($in);

echo "Total prompts: " . number_format($count) . "\n";

Converting to ChatML

def to_chatml(conv):
    return [
        {"role": "system",    "content": conv["system"]},
        {"role": "user",      "content": conv["prompt"]},
        {"role": "assistant", "content": conv["response"]},
    ]

Converting to Alpaca

def to_alpaca(conv):
    return {
        "instruction": conv["prompt"],
        "input":       "",
        "output":      conv["response"],
    }

---

📜 License

This project is licensed under the MIT License.

MIT License

Copyright (c) 2026 JibayAI

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.

✅ You Are Free To:

• Use this dataset for personal, academic, or commercial purposes.
• Modify and adapt it for your own pipelines and models.
• Redistribute it, in whole or in part, in any medium or format.
• Publish derivative works, including fine-tuned models and converted formats.

⚠️ The Only Requirement:

Attribution to "Jibay" must be preserved. Any redistribution, derivative work, or public use of this dataset must retain the name "Jibay" (and, where applicable, "JibayAI") in the license notice, README, or accompanying documentation. You may not remove, obscure, or replace the Jibay name with your own branding.

This is a non-negotiable condition of the MIT License as applied to this dataset. Removing or obscuring the Jibay attribution revokes your right to use, modify, or redistribute the dataset.

---

🙏 Acknowledgements

• JibayAI — for designing the Jibay Chat Format and curating this dataset.
• Open-source LLM community — for the tools, libraries, and inspiration that made this work possible.
• Persian NLP researchers — for their continued efforts to advance Persian-language AI.

Special thanks to the models that contributed to parts of this dataset: GPT-5 Codex, Claude Opus 4.6, and Jibay 4.1.

---

📬 Contact & Contributions

• Organization: JibayAI
• Format Spec: github.com/Jibay-Ai/Jibay_Chat_Format
• Issues & Pull Requests: Welcome via GitHub.

If you use this dataset in your research, product, or publication, please credit JibayAI and link back to the original repository.

---

⭐ Citation

If you use this dataset in academic work, please cite it as:

@misc{jibay_persian_chat_2026,
  title        = {Jibay Persian Chat Dataset},
  author       = {JibayAI},
  year         = {2026},
  howpublished = {\url{https://github.com/Jibay-Ai/Jibay_Chat_Format}},
  note         = {Persian-first instruction dataset in Jibay Chat Format (JCF). Licensed under MIT with mandatory Jibay attribution.}
}

---

Made with ❤️ by JibayAI
Persian-first. JCF-native. Open for everyone.