kailash-align

LLM fine-tuning and alignment framework for the Kailash platform. Part of the Kailash Python SDK by the Terrene Foundation.

Overview

kailash-align provides structured LoRA/QLoRA fine-tuning with full lifecycle tracking:

• AdapterRegistry -- Track adapters from training through evaluation, merge, GGUF export, and deployment
• AlignmentPipeline -- Thin wrappers around TRL SFTTrainer and DPOTrainer with checkpoint management
• AlignmentConfig -- Validated configuration dataclasses with NaN/Inf protection

Installation

pip install kailash-align

Optional Extras

pip install kailash-align[rlhf]    # QLoRA via bitsandbytes
pip install kailash-align[eval]    # Model evaluation via lm-eval
pip install kailash-align[serve]   # GGUF export via llama-cpp-python
pip install kailash-align[all]     # All extras

Quick Start

from kailash_align import AlignmentConfig, AlignmentPipeline, AdapterRegistry

# Configure
config = AlignmentConfig(
    base_model_id="meta-llama/Llama-3.1-8B",
    method="sft",
)

# Track adapters
registry = AdapterRegistry()

# Train
pipeline = AlignmentPipeline(config=config, adapter_registry=registry)
result = await pipeline.train(dataset=my_dataset, adapter_name="my-adapter")

# Result includes adapter path, metrics, and registry version
print(result.adapter_path)
print(result.training_metrics)

License

Apache 2.0 -- see LICENSE.