saara-ai 1.3.2

🧠 SAARA - Autonomous Document-to-LLM Data Engine. Pre-train, fine-tune, evaluate, and deploy LLMs.

🧠 SAARA: Autonomous Document-to-LLM Data Engine

SAARA is an end-to-end autonomous data pipeline designed to transform raw, unstructured documents (PDFs, research papers) into high-quality, instruction-tuned datasets for fine-tuning Large Language Models (LLMs).

Why this exists: Creating high-quality datasets is the bottleneck in training domain-specific AI. This tool automates the "boring stuff"—OCR, chunking, labeling, and cleaning—allowing you to go from PDF to fine-tuned model in hours, not weeks.

---

🚀 Key Features

1. 👁️ SOTA Vision-LLM OCR

• No more Garbled Text: Uses Moondream and Qwen2.5-VL (Vision-Language Models) to "read" PDFs visually.
• Handles complex double-column layouts, tables, and scientific diagrams that traditional OCR (Tesseract) fails on.
• Hybrid Fallback: Automatically switches between PyMuPDF (fast) and Vision OCR (accurate) based on page extractability.

2. 🤖 Autonomous Data Labeling

• Uses local LLMs (Granite 4.0, Llama 3) to generate diverse training tasks:
  • Instruction Tuning: "How do I treat X using Ayurveda?"
  • Q&A Pairs: Fact-based extraction.
  • Summarization: TL;DRs of complex sections.
  • Classification: Topic tagging.

3. 🧪 Data Distillation & Hygiene

• Self-Cleaning: The distill module removes low-quality generations, duplicates, and confabulations.
• ShareGPT Formatting: Automatically converts raw data into the industry-standard conversation format.

4. 🏗️ Pre-training from Scratch (NEW)

• Build Your Own LLM: Create custom models from 15M to 3B parameters.
• Custom Tokenizers: Train domain-specific BPE tokenizers on your data.
• Full Pipeline: Pre-train → Fine-tune → Evaluate → Deploy.
• Production-ready LLaMA-style architectures.

5. 🎓 Native Fine-Tuning Support

• One-Command Training: Built-in training loop using SFTTrainer (QLoRA) to fine-tune any HuggingFace model.
• Multi-Format Support: Automatically handles ShareGPT, Alpaca, and Raw Text formats.
• Checkpoint Resume: Continue training from any checkpoint.
• Iterative Fine-tuning: Fine-tune your already fine-tuned models to keep improving.
• Optimized for consumer GPUs (supports 4-bit quantization).

6. 🧪 Model Evaluation & Self-Improvement

• Granite 4 as Judge: Test your fine-tuned model with automatic quality scoring.
• Self-Improvement Loop: Low-scoring responses are corrected and used for next training round.
• Iterative Enhancement: Train → Evaluate → Improve → Repeat.

7. 🚀 Model Deployment

• Local Chat: Interactive terminal testing with your model.
• Ollama Export: Convert to GGUF format for Ollama usage.
• HuggingFace Hub: Push your model to share with the community.
• Cloud Deployment: Docker + Google Cloud Run ready.

---

🛠️ Architecture

graph LR
    A[Raw PDF] --> B(Vision OCR / Extractor)
    B --> C{Chunker Strategy}
    C --> D[Synthetic Labeling Agent]
    D --> E[Raw Dataset JSONL]
    E --> F(Data Distiller)
    F --> G[Clean ShareGPT Dataset]
    G --> H{Training Path}
    H -->|Pre-train| I[Build New Model]
    H -->|Fine-tune| J[Adapt Existing Model]
    I --> K[Model Evaluation]
    J --> K
    K --> L{Score < 7?}
    L -->|Yes| M[Generate Corrections]
    M --> J
    L -->|No| N((Deploy Model))

---

📦 Installation

1. Clone the repository:

   git clone https://github.com/nikhil49023/Data-engine.git
   cd Data-engine
   

2. Install the CLI:

   pip install -e .
   

3. Setup Ollama:

   • Install Ollama
   • The setup wizard will help you install models automatically

Quick Start

First-time setup (recommended):

saara setup

The setup wizard will:

1. ✅ Detect your hardware (GPU, VRAM, RAM)
2. ✅ Recommend optimal models for your system
3. ✅ Install selected vision and analyzer models
4. ✅ Save configuration

---

⚡ Usage

🎯 Interactive Wizard (Recommended)

saara run

This launches a beautiful CLI wizard with 5 workflows:

Option	Mode	Description
1	📄 Dataset Creation	Extract data from PDFs → Generate training datasets
2	🧠 Model Training	Fine-tune LLMs on your prepared data
3	🧪 Model Evaluation	Test & improve models with Granite 4
4	🚀 Model Deployment	Deploy locally (Ollama) or to cloud
5	🏗️ Pre-training	Build & train a model from scratch

---

🏗️ Pre-training from Scratch (NEW)

Build your own language model from the ground up:

saara pretrain

Available Architectures:

Name	Parameters	VRAM	Use Case
Nano	~15M	2GB+	Testing, learning (CPU trainable)
Micro	~50M	4GB+	Experimentation
Mini	~125M	6GB+	Domain-specific pre-training
Small	~350M	8GB+	Specialized tasks
Base	~1B	16GB+	Production models
Large	~3B	24GB+	High-capacity models

Pre-training Sub-menu:

1. 📚 Create Pre-training Dataset
2. 🏗️ Build & Train New Model
3. 🔤 Train Custom Tokenizer
4. 🧪 Test Pre-trained Model
5. 📋 List Pre-trained Models

Pre-training Dataset Creation:

• Extracts raw text from PDFs, markdown, and text files
• Cleans OCR artifacts and normalizes unicode
• Chunks text into optimal sizes for language modeling
• LLM-Enhanced Processing (Optional):
  • Uses local LLM (Granite 4, Llama 3, Qwen) to clean and improve text
  • Fixes OCR errors and expands abbreviations
  • LLM-based quality scoring for more accurate filtering
• Quality filtering (removes low-quality/incoherent text)
• Deduplication (prevents model memorization)
• Outputs in JSONL format ready for training
• Optional train/validation split

Workflow:

Create Dataset → Train Tokenizer (optional) → Pre-train Model → Test → Fine-tune → Deploy

---

📄 Dataset Creation Flow

1. Select input PDF folder and output directory
2. Choose Vision OCR model (Moondream/Qwen) - auto-detects available models
3. Choose Analyzer model (Granite 4/Llama 3/Qwen 2.5/Mistral)
4. Configure advanced options (chunk size, Q&A density)
5. Pipeline automatically generates:
   • *_instruction.jsonl - Instruction tuning data
   • *_qa.jsonl - Q&A pairs
   • *_sharegpt.jsonl - Chat format (best for training)
   • *_summarization.jsonl - Summarization tasks

---

🧠 Model Training Flow

The training wizard now supports:

• Base Models: HuggingFace models (Gemma, Llama, Qwen, etc.)
• Custom Pre-trained: Your own pre-trained models
• Fine-tuned Adapters: Continue training existing adapters

Supported Base Models:

Model	Size	Best For
sarvamai/sarvam-1	2B	Indian Languages
google/gemma-2b	2B	General Purpose
TinyLlama/TinyLlama-1.1B	1.1B	Fast Testing
meta-llama/Llama-3.2-1B	1B	English Tasks
Qwen/Qwen2.5-7B	7B	Complex Reasoning

Output: models/{model-name}-finetuned/final_adapter/

---

🧪 Model Evaluation Flow

Uses Granite 4 to evaluate your fine-tuned model:

1. Runs test prompts through your model
2. Scores each response (1-10)
3. Generates improved responses for low scores
4. Creates correction data for next training round

Self-Improvement Cycle:

Train Model → Evaluate (Granite 4) → Generate Corrections → Retrain → Repeat

---

🚀 Model Deployment Flow

Option	Platform	Description
1	Local Chat	Interactive terminal chat
2	Ollama Export	Convert to GGUF format
3	HuggingFace	Push to HF Hub
4	Cloud Deploy	Docker + Google Cloud Run
5	Merge Model	Merge adapter with base

---

📟 CLI Commands

Core Commands

Command	Description
saara run	Start interactive wizard
saara pretrain	Build & train model from scratch
saara setup	First-time hardware detection & model setup
saara version	Show version information

Data Processing

Command	Description
saara process <file>	Process a single PDF file
saara batch <dir>	Process all PDFs in directory
saara distill <input>	Generate synthetic training data

Model Operations

Command	Description
saara train	Fine-tune a model (interactive)
saara deploy	Deploy a trained model
saara evaluate <base> <adapter>	Evaluate model quality

Model Management

Command	Description
saara models list	List all available models
saara models install <name>	Install an Ollama model
saara models remove <name>	Remove a model
saara models status	Show hardware & model status

Server

Command	Description
saara serve	Start REST API server

---

📁 Project Structure

Data-engine/
├── setup.py                # Package setup
├── config.yaml             # Configuration settings
├── requirements.txt        # Dependencies
├── saara/                  # Source code
│   ├── cli.py             # CLI entry point
│   ├── pipeline.py         # Core data pipeline
│   ├── pretrain.py         # Pre-training module (NEW)
│   ├── train.py            # LLM fine-tuning module
│   ├── evaluator.py        # Model evaluation
│   ├── deployer.py         # Deployment utilities
│   ├── distiller.py        # Data cleaning
│   ├── model_manager.py    # Ollama model management
│   └── splash.py           # SAARA splash screen
├── models/                 # Saved models (pre-trained & fine-tuned)
├── datasets/               # Generated datasets
├── tokenizers/             # Custom tokenizers
├── evaluations/            # Evaluation results
└── exports/                # Deployment artifacts

---

🔮 Roadmap

• Vision-LLM OCR (Moondream, Qwen)
• Autonomous data labeling
• Multi-format dataset generation
• Native fine-tuning with QLoRA
• Model evaluation with Granite 4
• Self-improvement training loop
• Local & cloud deployment
• Pre-training from scratch
• Custom tokenizer training
• Iterative adapter fine-tuning
• Multi-modal dataset generation (images + text)
• RAG-based factual verification
• Web UI dashboard

---

📄 License

Proprietary License - Copyright © 2024-2025 Kilani Sai Nikhil. All Rights Reserved.

This software is provided under a proprietary license with the following terms:

✅ Permitted:

• Use the software for personal, educational, or commercial purposes
• Reference in academic/educational contexts with attribution

❌ Not Permitted:

• Modify, alter, or create derivative works
• Reproduce, copy, or duplicate the software
• Distribute, sublicense, or sell the software
• Reverse engineer or decompile the software

See the LICENSE file for full details.

---

👤 Author

Kilani Sai Nikhil - GitHub

---

Built with ❤️ for the AI community