TuneForge — LLM Fine-Tuning Kit

Overview

Prompting a hosted model is the right tool for most jobs — but for high-volume, format-strict, or domain-specific tasks, a fine-tuned model is cheaper per call, more consistent, and can run fully offline. TuneForge gives you that pipeline without having to wire together transformers, PEFT, and TRL yourself. You provide examples as simple JSONL (instruction, optional input, output); `prepare` renders them with the base model's chat template; `train` LoRA-fine-tunes an open base model, writing a small adapter to disk; and `serve` loads the base model plus your adapter behind a FastAPI endpoint so your app can call your own model like any API. Everything — the base model, LoRA rank, learning rate, sequence length — is controlled from one config.yaml, and the default base is a small open model so your first run works quickly before you scale up. This is a developer kit: training requires an NVIDIA GPU, and your data and the resulting model never leave your machine.

Key features

• Three-command pipeline: prepare → train → serve
• LoRA fine-tuning (transformers + PEFT + TRL) — trains on a single GPU
• Serve your fine-tuned model behind a FastAPI /generate endpoint
• Everything configured in one config.yaml — base model, LoRA, training args
• Swap in any open base model (Llama, Mistral, Qwen, …)
• Fully local — training data and the model never leave your machine
• Includes example training data and a CLI test tool

Specifications

Product type

Self-hosted developer kit (source code)

License

Perpetual, modify & redistribute in your products

Stack

Python · transformers · PEFT · TRL · FastAPI

Method

LoRA supervised fine-tuning

Default base model

Qwen2.5-0.5B-Instruct (swappable)

Requirements

NVIDIA GPU + CUDA for training (no API key)

Delivery

License key + private source download link