fattah-Orch-XS/README.md

---
tags:
- text-generation-inference
- transformers
- unsloth
license: apache-2.0
language:
- en
---
# Fattah-Orch — Arabic-First Coding Orchestrator

**Fattah-Orch** is a lightweight model that sits at the top of your AI coding pipeline. Give it a software request in **Egyptian Arabic, Modern Standard Arabic, or English** — it thinks through the requirements and returns a clean, structured JSON task graph your coding agents can execute directly.

<p align="center">
  <img src="https://huggingface.co/nomeda-lab/Fattah-Orch-Small/resolve/main/logo.png" alt="Fattah-Orch Logo" width="800"/>
</p>


## The Fattah-Orch Family

| Model | Parameters | Target Device |
|---|---|---|
| [Fattah-Orch-XS](https://huggingface.co/nomeda-lab/Fattah-Orch-XS) | 0.6B | Any CPU |
| [Fattah-Orch-S](https://huggingface.co/nomeda-lab/Fattah-Orch-Small) | 1.7B | CPU / Weak GPU |
| [Fattah-Orch-M](https://huggingface.co/nomeda-lab/Fattah-Orch-M) | 4B | GPU / Apple Silicon |
| [Fattah-Orch-L](https://huggingface.co/nomeda-lab/Fattah-Orch-L) | 8B | Mid GPU 8GB+ |

---

## What It Does

```
Your Request  →  Fattah-Orch  →  JSON Task Graph  →  Coder Model
(Arabic / English)   (local, fast)    (typed + ordered)    (GPT-4o, Claude, etc.)
```

Instead of sending a vague prompt directly to an expensive coder model, Fattah-Orch breaks it down into precise, typed, dependency-ordered subtasks first. The coder model gets clear instructions — no back-and-forth, fewer tokens, better output.

---

## Output Schema

```json
{
  "request_summary": "Full sentence describing what was requested",
  "subtasks": [
    {
      "id": 1,
      "title": "Short task name",
      "description": "What to build and what it should do",
      "type": "python",
      "depends_on": []
    },
    {
      "id": 2,
      "title": "Another task",
      "description": "What this builds and why it depends on task 1",
      "type": "typescript",
      "depends_on": [1]
    }
  ]
}
```

### Supported Task Types

| Type | When used |
|---|---|
| `python` | Backend, APIs, scripts |
| `typescript` | Frontend, React, Next.js |
| `sql` | Database schema, migrations |
| `go` | High-performance backend services |
| `kotlin` | Android native |
| `swift` | iOS native |
| `bash` | Shell scripts, infrastructure |

---

## Usage

### Installation

```bash
pip install unsloth transformers torch
```

### Inference

```python
import json
import torch
from unsloth import FastLanguageModel

MODEL_NAME = "nomeda-lab/Fattah-Orch-XS"  # or -S, -M, -L

SYSTEM_PROMPT = """You are Fattah-Orch, a software project orchestrator.

RULES:
1. Always include BOTH backend AND frontend tasks when the request implies a full system
2. Each subtask description must be 1-2 sentences explaining WHAT to build and WHAT it should do
3. request_summary must be a full sentence describing the complete system requested
4. Output ONLY valid JSON, nothing else

OUTPUT FORMAT:
{"request_summary": "...", "subtasks": [{"id": 1, "title": "...", "description": "...", "type": "python|typescript|sql|bash", "depends_on": []}]}"""

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=MODEL_NAME,
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

def orchestrate(request: str) -> dict:
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user",   "content": request},
    ]
    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
        enable_thinking=False,
    ).to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            input_ids=inputs,
            max_new_tokens=1024,
            temperature=0.3,
            top_p=0.9,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
        )

    response = tokenizer.decode(
        outputs[0][inputs.shape[-1]:],
        skip_special_tokens=True,
    )
    return json.loads(response)


# Arabic
plan = orchestrate("عايز تطبيق e-commerce فيه products و cart و checkout")
print(json.dumps(plan, indent=2, ensure_ascii=False))

# English
plan = orchestrate("I want a REST API for a blog with posts, comments and auth")
print(json.dumps(plan, indent=2, ensure_ascii=False))
```

---

## Example

**Input:** `"عايز تطبيق e-commerce فيه products و cart و checkout"`

```json
{
  "request_summary": "E-commerce application with product listing, shopping cart, and checkout flow",
  "subtasks": [
    {
      "id": 1,
      "title": "Product database model",
      "description": "Define Product model with name, price, stock, and category fields",
      "type": "python",
      "depends_on": []
    },
    {
      "id": 2,
      "title": "Products API",
      "description": "Endpoints to list, create, update, and delete products",
      "type": "python",
      "depends_on": [1]
    },
    {
      "id": 3,
      "title": "Cart and checkout API",
      "description": "Endpoints to add items to cart, view cart, and process checkout with order creation",
      "type": "python",
      "depends_on": [2]
    },
    {
      "id": 4,
      "title": "React storefront UI",
      "description": "Product listing page, cart sidebar, and checkout form that consumes the backend API",
      "type": "typescript",
      "depends_on": [2]
    }
  ]
}
```

---

## Limitations

- Best performance on Egyptian Arabic colloquial. MSA and other dialects work but may be less fluent.
- Task description quality improves with model size — XS is a fast baseline, L produces richer output.
- For very large systems (microservices, monorepos) prefer Orch-M or Orch-L.
- This model plans tasks — it does not write code. Connect it to a coder model for full end-to-end generation.

---

## Citation

```bibtex
@model{fattah_orch_2026,
  title  = {Fattah-Orch Family: Arabic-First Coding Orchestrator Models},
  author = {Nomeda Lab},
  year   = {2026},
  url    = {https://huggingface.co/collections/nomeda-lab/fattah-orch-family}
}
```

---

## License

[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — free for research and internal use; commercial redistribution requires permission.

---

*Part of the **Fattah project** — an open Arabic-first AI coding assistant ecosystem built at Nomeda Lab.*