README.md

# Haman Labs

### Open, graph-native infrastructure for Persian AI

Haman Labs is the AI research and open-source engineering lab of [Aria Haman Mehr Parseh](https://ariahaman.ir/en/). We build Persian-first infrastructure that transforms raw text into language graphs, graph-aware language models, reproducible datasets, and extensible AI products.

**Haman Labs is home to [Rakhshai Graph-based NLP (RGN)](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP), Iran’s first integrated graph-oriented NLP project for Persian, and [Haman Persian Article Graph-LLM 125M](https://huggingface.co/aria-haman/haman-fa-article-graph-llm-125m), the first Iranian LLM built with an [Iranian-produced architecture](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP).**

## What we build

### Rakhshai Graph-based NLP (RGN)

An open-source, graph-native platform for Persian natural language processing. RGN provides a coherent path from Persian text preparation and multi-relation graph construction to graph learning, language-model training, evaluation, inference, Python APIs, CLI tools, and MCP integration.

- [Source code](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP)
- [Documentation](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP/tree/main/docs)

### Haman Persian Article Graph-LLM 125M

A native Persian language model developed for structured article generation. The model accepts a topic, audience, tone, and section count and produces structured output in Markdown or JSON.

Its Iranian-produced architecture combines a decoder-only Transformer with a GCN over a corpus-level lexical graph and context-gated fusion between graph and token representations. Its tokenizer, graph, training workflow, and model weights were built for Persian, with training starting from random initialization.

- [Model card and weights](https://huggingface.co/aria-haman/haman-fa-article-graph-llm-125m)
- [Training dataset](https://huggingface.co/datasets/aria-haman/haman-fa-wikipedia-articles-186k)
- [Training notebook](https://colab.research.google.com/drive/1E50ISg1ANoW_rrFeRNBRcDn6C0cwfLyT?usp=sharing)

### Haman Persian Wikipedia Articles 186K

A dataset of approximately 186,000 Persian Wikipedia articles prepared to support reproducible Persian article-model training and experimentation.

## Our approach

- **Persian-first:** Built around Persian language, data, and linguistic needs.
- **Graph-native:** Relations, context, and structure are part of the modeling path.
- **Open and inspectable:** Source code, documentation, datasets, and workflows are available to the community.
- **Research to product:** A foundation for researchers, developers, AI agents, and production systems.

---

## دربارهٔ آزمایشگاه هامان

آزمایشگاه هوش مصنوعی هامان، واحد پژوهش و توسعهٔ متن‌باز شرکت دانش‌بنیان نرم‌افزاری آریا هامان مهر پارسه است. ما زیرساخت‌هایی می‌سازیم که رابطهٔ میان واژه‌ها، جمله‌ها و اسناد فارسی را به بخشی از هوشمندی مدل تبدیل می‌کنند.

**آزمایشگاه هامان خانهٔ ⁦[Rakhshai Graph-based NLP (RGN)](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP)⁩، نخستین پروژهٔ ایرانی یکپارچهٔ ⁦NLP⁩ گراف‌محور برای زبان فارسی، و مدل ⁦[Haman Persian Article Graph-LLM 125M](https://huggingface.co/aria-haman/haman-fa-article-graph-llm-125m)⁩، نخستین ⁦LLM⁩ ایرانی ساخته‌شده با [معماری تولید ایرانی](https://github.com/bazpardazesh-org/Rakhshai-Graph-based-NLP)، است.**

هدف ما توسعهٔ فناوری فارسی‌محور، دانش باز و ابزارهایی است که مسیر میان پژوهش، مدل زبانی و محصول را برای جامعهٔ فارسی‌زبان کوتاه‌تر می‌کنند.

## Links

- [Haman Labs](https://hamanlabs.ir/)
- [GitHub](https://github.com/bazpardazesh-org)
- [Hugging Face model](https://huggingface.co/aria-haman/haman-fa-article-graph-llm-125m)
- [Hugging Face dataset](https://huggingface.co/datasets/aria-haman/haman-fa-wikipedia-articles-186k)
- [Aria Haman Mehr Parseh](https://ariahaman.ir/en/)