Skip to content

Fine-Tune Pipeline

Welcome to the Fine-Tune Pipeline documentation! This is a comprehensive fine-tuning pipeline for language models that includes training, inference, and evaluation capabilities.

🚀 Features

  • Easy Configuration: TOML-based configuration system for all components
  • Modern Architecture: Built with Unsloth, Transformers, and TRL for efficient fine-tuning
  • Comprehensive Evaluation: Multiple evaluation metrics including Factual Correctness, Answer Accuracy, and more
  • CI/CD Ready: Designed to work seamlessly with GitHub Actions and other CI/CD tools
  • Flexible Inference: Built-in inference capabilities with customizable parameters
  • MLflow Integration: Comprehensive experiment tracking, model versioning, comparing, and more
  • Pipeline Orchestration: Integrated pipeline runner with phase control

🏗️ Architecture

The pipeline consists of following components:

1. Pipeline Orchestrator (app/pipeline_invoker.py)

  • Coordinates the execution of fine-tuning, inference, and evaluation phases
  • Comprehensive MLflow experiment tracking and logging
  • Configurable phase execution with stop-after options

2. Fine-Tuner (app/finetuner.py)

  • Handles model fine-tuning using LoRA (Low-Rank Adaptation)
  • Supports 4-bit and 8-bit quantization for memory efficiency
  • Integrates with MLFlow for experiment tracking
  • Automatic model publishing to Hugging Face Hub

3. Inferencer (app/inferencer.py)

  • Performs inference on test datasets
  • Configurable generation parameters
  • Outputs results in JSONL format
  • Pushes results to Hugging Face Hub

4. Evaluator (app/evaluator.py)

  • Comprehensive evaluation suite with multiple metrics
  • Support for both traditional (BLEU, ROUGE) and LLM-based evaluation
  • Detailed reporting with Excel and JSON outputs

5. Documentation Server (app/docs_server.py)

  • Built-in documentation server with MkDocs integration
  • Static documentation build capabilities
  • Serve documentation in GitHub Pages

🛠️ Technology Stack

  • Unsloth: Efficient fine-tuning framework
  • Transformers: Hugging Face Transformers library
  • TRL: Transformer Reinforcement Learning
  • Datasets: Hugging Face Datasets library
  • MLflow: Machine learning lifecycle management and experiment tracking
  • RAGAS: Retrieval Augmented Generation Assessment
  • MkDocs: Documentation generator

📊 Supported Models

The pipeline supports various model architectures including:

  • Qwen 2.5 series
  • Llama models
  • Gemma models
  • And any model compatible with Unsloth

📈 Evaluation Metrics

Built-in support for:

  • BLEU Score: Translation quality assessment
  • ROUGE Score: Summarization evaluation
  • Factual Correctness: LLM-based factual evaluation
  • Semantic Similarity: Embedding-based similarity
  • Answer Accuracy: Model response and ground truth agreement
  • Answer Relevancy: Relevance assessment

🔧 Quick Start

Ready to get started? Check out our Quick Start Guide to begin fine-tuning your first model!

📚 What's Next?

For issues or contributions, please visit our GitHub repository.