Fine-tune a model on your own data
Pick a base model, point it at one of your datasets, and train with LoRA on Inferix GPUs. No training loop to write, no CUDA environment to keep alive, and no machine to leave running when the job finishes.
Bring your own dataset
Train on any dataset you can reach on Inferix. Columns are read straight from the dataset, so you choose which field is the text to learn from.
LoRA, not full fine-tuning
Low-rank adapters train a small fraction of the weights. Jobs finish faster, cost less GPU time, and the result is a small adapter rather than a full model copy.
Watch it run
Jobs report status while they train and can be cancelled mid-run. Nothing keeps billing after you stop it.
Base models available today
Training currently supports four small language models. They are chosen to fine-tune quickly and cheaply rather than to top a benchmark — useful for style, formatting and domain vocabulary, not for teaching a model new capabilities.
GPT-2
gpt2 · 124M parameters
DistilGPT-2
distilgpt2 · 82M parameters
Pythia 160M
EleutherAI/pythia-160m · 160M parameters
Pythia 410M
EleutherAI/pythia-410m · 410M parameters
Larger base models are not offered yet. When they are, they will appear here rather than being promised in advance.
Sign in to start a job, or read the getting-started guide first.