aligntune 0.1.7

AlignTune: Multi-backend alignment and fine-tuning library. Features TRL and Unsloth backends with complete RL coverage (DPO, PPO, GRPO, BOLT), 27+ reward functions, production-ready evaluation system, and unified configuration interface for LLM alignment.

---

AlignTune is a production-ready fine-tuning library designed to simplify training and fine-tuning of Large Language Models (LLMs) with both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) methods. It provides a high-level, unified API that abstracts away the complexities of backend selection, algorithm configuration, and training loops, letting you focus on delivering results.

Core Features

Multi-Backend Architecture: Choose between TRL (reliable, battle-tested) and Unsloth (faster) backends with intelligent auto-selection.

Complete RLHF Coverage: 12+ RL algorithms including DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, and PACE.

Production-Ready: No mock code, comprehensive error handling, extensive testing, and robust validation.

Quick Start

Supervised Fine-Tuning (SFT)

from aligntune.core.backend_factory import create_sft_trainer

# Create and train SFT model
trainer = create_sft_trainer(
    model_name="microsoft/DialoGPT-small",
    dataset_name="tatsu-lab/alpaca",
    backend="trl",
    num_epochs=3,
    max_steps = -1,
    batch_size=4,
    learning_rate=5e-5
)

# Train the model
trainer.train()

# Evaluate
metrics = trainer.evaluate()
print(metrics)

Reinforcement Learning (DPO)

from aligntune.core.backend_factory import create_rl_trainer

# Create and train DPO model
trainer = create_rl_trainer(
    model_name="Qwen/Qwen3-0.6B",
    dataset_name="Anthropic/hh-rlhf",
    algorithm="dpo",
    backend="trl",
    num_epochs=1,
    batch_size=4,
    learning_rate=5e-5
)

# Train the model
trainer.train()

Supported Algorithms

Algorithm	TRL Backend	Unsloth Backend	Description
SFT	Yes	Yes	Supervised Fine-Tuning
DPO	Yes	Yes	Direct Preference Optimization
PPO	Yes	Yes	Proximal Policy Optimization
GRPO	Yes	Yes	Group Relative Policy Optimization
GSPO	Yes	Yes	Group Sequential Policy Optimization
DAPO	Yes	Yes	Decouple Clip and Dynamic sAmpling Policy Optimization
Dr. GRPO	Yes	Yes	GRPO Done Right (unbiased variant)
GBMPO	Yes	No	Group-Based Mirror Policy Optimization
Counterfactual GRPO	Yes	Yes	Counterfactual GRPO variant
PACE	Yes	Yes	Baseline-Optimized Learning Technique

Installation

# Or install from source
git clone https://github.com/Lexsi-Labs/aligntune.git
cd aligntune
pip install -e .

Requirements

• Python 3.12+
• PyTorch 2.0+
• CUDA-compatible GPU (recommended for faster training)

Demo Notebooks

Interactive Colab notebooks demonstrating various AlignTune workflows: Here are the organized tables containing the Colab links, models, and datasets provided in your text.

Supervised Fine-Tuning (SFT)

Backend	Model	Dataset	Link
TRL	Qwen/Qwen3-4B-Instruct-2507	sohamb37lexsi/bitext-wealth-management-llm-chatbot-splits
TRL	Qwen3-4B-Instruct	sohamb37lexsi/bitext-retail-banking-llm-chatbot-splits
Unsloth	Qwen/Qwen2.5-0.5B-Instruct	bebechien/MobileGameNPC
TRL	google/txgemma-2b-predict	trialbench_adverse-event-rate-prediction
Unsloth	Qwen/Qwen2.5-0.5B-Instruct	bebechien/MobileGameNP

Reinforcement Learning (RL)

Backend	Algorithm	Model	Dataset	Link
Unsloth	DPO	microsoft/phi-2	argilla/distilabel-intel-orca-dpo-pairs
TRL	DPO	google/gemma-2-2b-it	Anthropic/hh-rlhf
TRL	DPO	sohamb37lexsi/wealth_management_Qwen3-4B-Instruct-2507	sohamb37lexsi/bitext_wealth_management_preference_data
Unsloth	PPO	Qwen/Qwen2.5-0.5B-Instruct	HuggingFaceH4/ultrachat_200k
TRL	PPO	EleutherAI/pythia-1.4b	CarperAI/openai_summarize_tldr
TRL	GRPO (Coding)	Qwen/Qwen3-4B	google-research-datasets/mbpp
Unsloth	GRPO (Math)	meta-llama/Llama-3.2-3B-Instruct	openai/gsm8k
TRL	GRPO	meta-llama/Llama-3.2-3B-Instruct	openai/gsm8k
Unsloth	DRGRPO	Qwen/Qwen2.5-3B-Instruct	yahma/alpaca-cleaned
TRL	DRGRPO	Qwen/Qwen2-0.5B-Instruct	AI-MO/NuminaMath-TIR
Unsloth	GSPO	Qwen/Qwen3-1.7B	CyberNative/Code_Vulnerability_Security_DPO
TRL	GSPO	meta-llama/Llama-3.2-3B-Instruct	HuggingFaceH4/ultrachat_200k
Unsloth	DAPO	microsoft/Phi-3.5-mini-instruct	HuggingFaceH4/ultrachat_200k
TRL	DAPO	meta-llama/Llama-3.2-3B-Instruct	google-research-datasets/mbpp

Documentation

• Getting Started: Installation, setup, and basic usage
• User Guide: In-depth tutorials for SFT and RL training
• API Reference: Complete Python API and class/method details
• Examples: End-to-end code examples
• Advanced Topics: Architecture, custom backends, and performance optimization
• Notebooks: Interactive Colab notebooks and local Jupyter notebooks

Key Capabilities

• Multiple Training Paradigms: Supports SFT, DPO, PPO, GRPO, and advanced RL algorithms
• Backend Flexibility: TRL and Unsloth backends with automatic fallback
• Reward Model Training: Train custom reward models from rule-based functions
• Comprehensive Evaluation: Multi-level evaluation with lm-eval integration
• Production Ready: Model serialization, reproducible training, and deployment-ready pipelines
• Extensible Architecture: Modular design for easy integration of custom algorithms and backends

Architecture

AlignTune uses a flexible backend architecture:

flowchart TD
    Factory[Backend Factory] --> TRL[TRL Backend]
    Factory --> Unsloth[Unsloth Backend]
    TRL --> TRL_Algos[TRL Algorithms]
    Unsloth --> Unsloth_Algos[Unsloth Algorithms]

TRL Backend: SFT, DPO, PPO, GRPO, GSPO, DAPO, Dr. GRPO, GBMPO, Counterfactual GRPO, PACE

Unsloth Backend: SFT, DPO, PPO, GRPO, DAPO, Dr. GRPO, Counterfactual GRPO, PACE

See Architecture for details.

Contributing

We welcome contributions! See our Contributing Guide for details.

License

This project is released under the MIT License. Please cite appropriately if used in academic or production projects. See the LICENSE file for details.

Key Points:

• Free for Research & Learning: Use, modify, and study for personal, academic, or research purposes
• Source Available: Full access to source code
• Commercial Use Restricted: Requires separate commercial license
• Contact: For commercial licensing, partnership, or redistribution rights, contact [e-mail]

This is not an open-source license as defined by OSI, but provides broad access for non-commercial use.

Citation

If you use AlignTune in your research, please cite:

BibTeX:

@software{alignTune2025,
  title        = {{AlignTune}: Modular Toolkit for Post-Training Alignment of Large Language Models},
  author       = {Lyngkhoi, R E Zera Marveen and Chawla, Chirag and Seth, Pratinav and Avaiya, Utsav and Bhattacharjee, Soham and Khandoga, Mykola and Yuan, Rui and Sankarapu, Vinay Kumar},
  year         = {2025},
  note         = {Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth},
  organization = {Lexsi Labs},
  url          = {https://github.com/Lexsi-Labs/aligntune},
  version      = {0.0.0}
}

Plain Text:

Lyngkhoi, R. E. Z. M., Chawla, C., Seth, P., Avaiya, U., Bhattacharjee, S., Khandoga, M.,
Yuan, R., & Sankarapu, V. K. (2025). AlignTune: Modular Toolkit for Post-Training Alignment
of Large Language Models. Lexsi Labs. https://github.com/Lexsi-Labs/aligntune

*Equal contribution: R E Zera Marveen Lyngkhoi, Chirag Chawla, Pratinav Seth

Acknowledgments

AlignTune is built upon the excellent work of the following projects:

• HuggingFace Transformers - Model architectures and tokenizers
• TRL - Transformer Reinforcement Learning library
• Unsloth - Fast and memory-efficient training
• HuggingFace Datasets - Dataset loading and processing

Support

• Documentation: aligntune.lexsi.ai/
• GitHub Issues: github.com/Lexsi-Labs/aligntune/issues
• Discussions: github.com/Lexsi-Labs/aligntune/discussions
• Email: [e-mail]
• Discord: Discord Lexsi Labs

---

Get started with AlignTune and accelerate your LLM fine-tuning workflows today!

Contact

https://www.lexsi.ai

Paris 🇫🇷 · Mumbai 🇮🇳 · London 🇬🇧