kailash-align

What problem does it solve?

It prevents unsafe or underperforming fine-tuned LLM deployments by providing an end-to-end training, evaluation, and serving workflow with a hard evaluation-before-serving gate.

Core Features & Use Cases

• End-to-end alignment pipeline: Orchestrates training across multiple TRL-based alignment methods and adapter lifecycles.
• 12 supported alignment methods: Covers offline, unpaired, monolithic, and online training patterns (including LoRA-based adapter management and chaining for multi-stage flows like sft_then_dpo).
• Mandatory eval-before-serve: Requires benchmarking and safety checks (including base-model comparison) before any model can be deployed.
• Deployment options for real runtimes: Exports to GGUF, deploys to Ollama, serves via vLLM, and connects deployed models into Kaizen agents through a bridge.

Quick Start

Ask the AI to train and align a DPO adapter, then evaluate it and deploy it to Ollama or vLLM only if evaluation passes.