Tunedmodel

Opinionated, end-to-end framework for shipping fine-tuned language models.

tunedmodel.com

What is Tunedmodel?

Tunedmodel is an Elixir-first framework for building, evaluating, and shipping fine-tuned language models.

It provides a consistent project shape and tooling so a “model” is more than weights:

• Training recipes you can rerun
• Evals that gate releases
• Versioned artifacts and configuration
• A runtime you can run locally or deploy

What is a Tunedmodel Project?

A Tunedmodel project is the unit you build and ship.

It’s a repository that bundles everything required to go from:

• data → fine-tune → eval → release bundle

In practice, a project typically includes:

• Data sources/connectors
• Fine-tuning steps and parameters
• Evaluation datasets and scoring
• Runtime defaults (limits, policies, routing)
• Optional chat UX and tool integrations

Key capabilities (planned)

• Data connectors for files, repos, and URLs
• Repeatable training recipes that produce release bundles
• Built-in testing via tunedmodel test to run evals and gate releases
• Skill packs for common capabilities
• Built-in coding assistant via tunedmodel code

Typical workflow (planned)

• Run tunedmodel new to scaffold a project
• Add data through connectors
• Run tunedmodel build to produce a release bundle
• Run tunedmodel start to run the runtime

Tech stack (Elixir)

Tunedmodel is centered on Elixir/OTP, with likely integrations across:

• Phoenix / LiveView for the runtime UI
• Nx / Axon / Bumblebee for model execution
• Oban / Broadway for durable pipelines

Installation

If/when Tunedmodel is published to Hex, install by adding tunedmodel to mix.exs:

def deps do
  [
    {:tunedmodel, "~> 0.1.1"}
  ]
end

HexDocs: https://hexdocs.pm/tunedmodel/