training-hub 0.8.0

An algorithm-focused interface for common language model training, continual learning, and reinforcement learning techniques

Training Hub

Training Hub is an algorithm-focused interface for common LLM training, continual learning, and reinforcement learning techniques developed by the Red Hat AI Innovation Team.

New to Training Hub? Read our comprehensive introduction: Get Started with Language Model Post-Training Using Training Hub

Support Matrix

Algorithm	InstructLab-Training	RHAI Innovation Mini-Trainer	PEFT	Unsloth	verl	Status
Supervised Fine-tuning (SFT)	✅	-	-	-	-	Implemented
Continual Learning (OSFT)	🔄	✅	🔄	-	-	Implemented
Low-Rank Adaptation (LoRA) + SFT	-	-	-	✅	-	Implemented
LoRA + GRPO (Adapter-Based RLVR)	-	-	-	✅	✅	Implemented
Direct Preference Optimization (DPO)	-	-	-	-	🔄	Planned

Legend:

• ✅ Implemented and tested
• 🔄 Planned for future implementation
• - Not applicable or not planned

Implemented Algorithms

Supervised Fine-tuning (SFT)

Fine-tune language models on supervised datasets with support for:

• Single-node and multi-node distributed training
• Configurable training parameters (epochs, batch size, learning rate, etc.)
• InstructLab-Training backend integration

from training_hub import sft

result = sft(
    model_path="Qwen/Qwen2.5-1.5B-Instruct",
    data_path="/path/to/data",
    ckpt_output_dir="/path/to/checkpoints",
    num_epochs=3,
    effective_batch_size=8,
    learning_rate=1e-5,
    max_seq_len=256,
    max_tokens_per_gpu=1024,
)

Orthogonal Subspace Fine-Tuning (OSFT)

OSFT allows you to fine-tune models while controlling how much of its existing behavior to preserve. Currently we have support for:

• Single-node and multi-node distributed training
• Configurable training parameters (epochs, batch size, learning rate, etc.)
• RHAI Innovation Mini-Trainer backend integration

Here's a quick and minimal way to get started with OSFT:

from training_hub import osft

result = osft(
    model_path="/path/to/model",
    data_path="/path/to/data.jsonl", 
    ckpt_output_dir="/path/to/outputs",
    unfreeze_rank_ratio=0.25,
    effective_batch_size=16,
    max_tokens_per_gpu=2048,
    max_seq_len=1024,
    learning_rate=5e-6,
)

Low-Rank Adaptation (LoRA) + SFT

Parameter-efficient fine-tuning using LoRA with supervised fine-tuning. Features:

• Memory-efficient training with significantly reduced VRAM requirements
• Single-GPU and multi-GPU distributed training support
• Unsloth backend for 2x faster training and 70% less memory usage
• Support for QLoRA (4-bit quantization) for even lower memory usage
• Compatible with messages and Alpaca dataset formats

from training_hub import lora_sft

result = lora_sft(
    model_path="Qwen/Qwen2.5-1.5B-Instruct",
    data_path="/path/to/data.jsonl",
    ckpt_output_dir="/path/to/outputs",
    lora_r=16,
    lora_alpha=32,
    num_epochs=3,
    learning_rate=2e-4
)

LoRA + GRPO (Adapter-Based RLVR)

Train LoRA adapters on tool-calling agents using Group Relative Policy Optimization with reinforcement learning from verifiable rewards. Features:

• Single-turn and multi-turn tool-call verification with automatic per-turn decomposition
• Two backends: OpenPipe ART + Unsloth GRPO (single-GPU, fast iteration) and verl (multi-GPU, scales to 70B+)
• Built-in reward functions for tool-call correctness, or bring your own
• Zero API cost training using ground-truth trace decomposition

from training_hub import lora_grpo

# Single GPU (ART backend)
result = lora_grpo(
    model_path="Qwen/Qwen3-4B",
    data_path="./tool_call_traces.jsonl",
    ckpt_output_dir="./grpo_output",
    backend="art",
    lora_r=32,
    lora_alpha=64,
    num_iterations=15,
)

# Multi GPU (verl backend)
result = lora_grpo(
    model_path="Qwen/Qwen3-4B",
    data_path="./tool_call_traces.jsonl",
    ckpt_output_dir="./grpo_output",
    backend="verl",
    n_gpus=4,
)

Installation

Basic Installation

This installs the base package, but doesn't install the CUDA-related dependencies which are required for GPU training.

pip install training-hub

Development Installation

git clone https://github.com/Red-Hat-AI-Innovation-Team/training_hub
cd training_hub
pip install -e .

For developers: See the Development Guide for detailed instructions on setting up your development environment, running local documentation, and contributing to Training Hub.

LoRA Support

For LoRA training with optimized dependencies:

pip install training-hub[lora]
# or for development
pip install -e .[lora]

Note: The LoRA extras include Unsloth optimizations and PyTorch-optimized xformers for better performance and compatibility.

GRPO Support

For LoRA + GRPO training (both ART and verl backends):

pip install training-hub[grpo,lora]

Note: When combining [grpo] with [cuda] extras, install them sequentially to avoid dependency solver conflicts:

pip install training-hub[grpo,lora]
pip install training-hub[cuda]

The [grpo] extras constrain torch, vllm, and transformers versions for verl compatibility, which may conflict with versions pulled by [cuda]. Sequential installation lets the solver pick compatible versions.

CUDA Support

For GPU training with CUDA support:

pip install training-hub[cuda] --no-build-isolation
# or for development
pip install -e .[cuda] --no-build-isolation

Note: If you encounter build issues with flash-attn, install the base package first:

# Install base package (provides torch, packaging, wheel, ninja)
pip install training-hub
# Then install with CUDA extras
pip install training-hub[cuda] --no-build-isolation

# For development installation:
pip install -e . && pip install -e .[cuda] --no-build-isolation

If you're using uv, you can use the following commands to install the package:

# Installs training-hub from PyPI
uv pip install training-hub && uv pip install training-hub[cuda] --no-build-isolation

# For development:
git clone https://github.com/Red-Hat-AI-Innovation-Team/training_hub
cd training_hub
uv pip install -e . && uv pip install -e .[cuda] --no-build-isolation

Getting Started

For comprehensive tutorials, examples, and documentation, see the examples directory.