Unsloth + Ertas

Overview

Unsloth is a fast open-source fine-tuning library, delivering large training speed improvements through custom CUDA kernels and memory-efficient implementations. ML engineers reach for it because of its raw performance and minimal abstraction: you write Python in a notebook or script, call its patched model loader, and get quicker LoRA and QLoRA training at lower VRAM. For engineers who live in notebooks and want low-level control, that is a strong trade.

It is deliberately scoped to the training step. Dataset preparation, run history, quantisation and export, and getting the finished model onto a machine are all left to you. Teams using it assemble those stages from separate tools: a tracking service, custom data scripts, llama.cpp for quantisation, and manual steps for deployment. That assembly is fine for one engineer running experiments and becomes the bottleneck once a team needs reproducibility or a repeatable path from data to a shipped model.

How Ertas Integrates

Ertas takes the opposite approach: it is a managed pipeline where the whole job lives in one place and no Python is required. You prepare data in Data Craft, drop an Action Module on the Model Studio canvas, connect a base model, dataset and training config, and press play. Training runs on managed cloud GPUs, so there is no environment to build and no VRAM budget to fit into.

The practical differences show up on either side of the training step. Data Craft holds your datasets and Prompt Studio helps you author a generation prompt when you are short of real examples. Every run keeps its config snapshot and metrics, so you can open a model months later and see exactly what produced it. Finished runs export a LoRA adapter and a GGUF bundle that includes a pre-built Ollama Modelfile carrying the right chat template and stop tokens, plus install scripts, which is the step that most often costs an afternoon when you assemble the stack yourself.

Where Unsloth stays ahead is control. Custom training loops, unusual architectures, and hand-tuned kernel-level work are what a library gives you and a managed pipeline does not. If that is the job, a notebook is the right tool.

Getting Started

1. 1

   Prepare your dataset in Data Craft

   Upload JSONL, CSV, parquet or text, or mirror a dataset in from Hugging Face. When you are short of real examples, Prompt Studio builds a generation prompt from your own context that you run in the AI tool you already pay for and import back as JSONL.

2. 2

   Build the run on the canvas

   Drop an Action Module into a Model Studio project and connect its four legs: base model, dataset, training config and LoRA config. The canvas shows which legs are still missing, so an incomplete run cannot be submitted by accident.

3. 3

   Train on managed GPUs, in parallel

   Press play. Training runs on cloud GPUs with no environment setup. Each Action Module is an independent job and nothing caps how many run at once, so you can bracket hyperparameters across several modules simultaneously. Your credit balance is the practical limit.

4. 4

   Check the run and compare against earlier ones

   A finished run shows three inference samples from a fixed prompt set as a first coherence check, alongside its full config snapshot and metrics. Earlier runs in the project stay available for comparison, so you can see what changed between attempts.

5. 5

   Export and install

   Download the LoRA adapter as safetensors, or the GGUF bundle: a Q4_K_M quantised model file, a pre-built Modelfile with the chat template, stop tokens and your configured sampling defaults, and install scripts for macOS, Linux and Windows. Run the installer and Ollama serves the model.

Benefits

• No Python, no notebook, and no CUDA environment to build before the first training run
• Managed cloud GPUs, so model size is not bounded by the VRAM on your machine
• Data Craft and Prompt Studio cover dataset preparation in the same place you train
• Parallel training across Action Modules with no hard concurrency cap on any plan
• Every run keeps its config snapshot and metrics, so a model traces back to what produced it
• GGUF export ships as an installable bundle with a pre-built Ollama Modelfile and install scripts
• Usable by teammates who would not otherwise write a training script