saara-ai 0.1.1

End-to-end ML pipeline: PDF→JSON datasets, synthetic data, fine-tuning, evaluation, and edge deployment with Gradio GUI

SAARA

SAARA is a Python package for building local LLM workflows around data creation, fine-tuning, evaluation, export, and visualization.

It is aimed at a practical teacher-student workflow:

1. extract text from PDFs or raw text
2. generate synthetic supervision with a stronger local model
3. fine-tune a smaller model
4. evaluate quality, speed, memory, and optional teacher agreement
5. export and quantize for deployment
6. inspect the workflow through CLI or Gradio

Features

• PDF to text extraction with PyMuPDF and pdfplumber
• synthetic dataset generation for factual, reasoning, conversational, instruction, code, and creative tasks
• local provider abstraction for Ollama, vLLM, and llama.cpp
• fine-tuning helpers for LoRA and QLoRA workflows
• evaluation helpers for custom datasets, benchmarks, and teacher-student comparison
• export helpers for safetensors, GGUF, AWQ, GPTQ, ONNX, TensorRT, and PyTorch formats
• Gradio dashboard and CLI entrypoints

Install

From PyPI:

pip install saara-ai

Common installs:

pip install "saara-ai[training]"
pip install "saara-ai[export]"
pip install "saara-ai[providers,training,evaluation,export]"

From source:

git clone https://github.com/nikhil49023/saara-ai.git
cd saara-ai
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Optional extras:

pip install -e ".[dev]"
pip install -e ".[edge]"
pip install -e ".[providers,training,evaluation,export]"

Base saara-ai intentionally keeps the install lighter. Heavy native stacks such as auto-gptq, autoawq, llama-cpp-python, vllm, and training toolchains are opt-in extras.

Note: a virtual environment is strongly recommended.

Quick Start

1. Start a local model provider

For Ollama:

ollama pull mistral
ollama serve

2. Build a dataset from a PDF

from saara import DatasetBuilder
from saara.dataset.types import DataType
from saara.providers.ollama_provider import OllamaProvider, ProviderConfig

provider = OllamaProvider(
    ProviderConfig(model="mistral", base_url="http://localhost:11434")
)

builder = DatasetBuilder(provider)
samples = builder.from_pdf(
    "document.pdf",
    data_types=[DataType.INSTRUCTION, DataType.FACTUAL],
    pairs_per_type=5,
    min_quality=0.65,
)

builder.save(samples, "dataset.jsonl")

3. Launch the GUI

saara gui

Core Workflow

Dataset creation

from saara import DatasetBuilder
from saara.dataset.types import DataType

builder = DatasetBuilder(provider)
samples = builder.from_text(
    "Transformers use attention mechanisms for sequence modeling.",
    data_types=[DataType.FACTUAL, DataType.INSTRUCTION],
    pairs_per_type=3,
)

Fine-tuning

from saara import FineTuner
from saara.training.config import TrainingConfig

config = TrainingConfig(
    model_name="mistralai/Mistral-7B-v0.1",
    num_train_epochs=3,
    use_lora=True,
)

finetuner = FineTuner(config)
finetuner.train("dataset.jsonl")
finetuner.save("./output/models/my-finetune")

Evaluation

from saara import ModelEvaluator
from saara.providers.ollama_provider import OllamaProvider, ProviderConfig

student = OllamaProvider(ProviderConfig(model="mistral-7b-finetuned"))
teacher = OllamaProvider(ProviderConfig(model="llama-3-70b"))

evaluator = ModelEvaluator(student, teacher)
metrics = evaluator.evaluate(
    "test.jsonl",
    run_benchmarks=True,
    benchmark_names=["mmlu", "gsm8k"],
)

print(metrics.summary())

Export and quantization

from saara import ModelExporter
from saara.export.formats import ExportFormat
from saara.export.quantization import QuantizationConfig

exporter = ModelExporter(
    model="./output/models/my-finetune",
    config=QuantizationConfig(bits=4),
)

results = exporter.export(
    "./output/exports",
    formats=[ExportFormat.GGUF, ExportFormat.AWQ, ExportFormat.ONNX],
    quantize=True,
)

CLI

saara dataset from-pdf document.pdf -o ./output --data-types instruction factual
saara train finetune dataset.jsonl --model mistralai/Mistral-7B-v0.1 --epochs 3
saara eval model test.jsonl --model mistral --teacher llama-3-70b --benchmarks mmlu gsm8k
saara export model ./models/final --formats gguf awq --quantize --bits 4
saara gui --port 7860

Documentation

• User guide: USER_GUIDE.md
• Package structure: PACKAGE_STRUCTURE.md
• Examples: examples/

Examples

• examples/01_pdf_to_dataset.py
• examples/02_synthetic_data.py
• examples/03_finetune.py
• examples/04_evaluate.py
• examples/05_export.py
• examples/06_complete_pipeline.py
• examples/07_gui.py

Recommended Starting Point

If you're trying SAARA for the first time, start with:

• OllamaProvider
• one small PDF or a short text file
• LoRA or QLoRA fine-tuning
• GGUF and safetensors export first

Development

pytest tests/
black saara/
ruff check saara/
mypy saara/

PyPI

https://pypi.org/project/saara-ai/

License

MIT. See LICENSE.