Solutions
·Fine-Tuning
Adapt any model
to your data.
Run LoRA, QLoRA, or full-parameter fine-tuning on bare-metal GPUs. Hugging Face Transformers, Axolotl, and LLaMA-Factory come pre-installed. Bring your dataset, pick a base model, and start adapting -- per-minute billing means you only pay for active training time.
The Workflow
Every fine-tuning method, one platform.
LoRA & QLoRA
Train lightweight adapters on top of frozen base models. QLoRA with 4-bit quantization lets you fine-tune 70B models on a single GPU.
Full-parameter training
Unfreeze all layers for maximum performance. Multi-GPU instances with NVLink for models that need full fine-tuning.
Pre-installed frameworks
Hugging Face Transformers, Axolotl, LLaMA-Factory, PEFT, DeepSpeed, and bitsandbytes ready out of the box. No setup required.
Dataset management
Upload datasets directly to your instance via SCP, or pull from Hugging Face Hub. Persistent storage across sessions so you never re-upload.
Experiment tracking
Weights & Biases, MLflow, and TensorBoard all work out of the box. Track loss curves, hyperparameters, and checkpoints across runs.
Deploy when done
Export your adapter or merged model directly to the Runcrate Inference API or a dedicated serving instance. No data transfer needed.
Recommended Setups
Match your method
to the right hardware.
LoRA on a single L40S or full fine-tuning across H100s -- pick the setup that fits your model size and budget.
QLoRA on 7-13B modelsL40S · 48 GBFrom $0.50/hrLoRA on 70B modelsA100 · 80 GBFrom $1.10/hrFull fine-tune up to 70BH100 · 80 GBFrom $2.49/hrFull fine-tune 100B+H200 · 141 GBFrom $3.49/hrHow It Works
Three steps to a custom model.
Pick a base model and method
Choose any open-source model from Hugging Face. Select LoRA, QLoRA, or full fine-tuning based on your model size and dataset.
Upload data and configure
Upload your training data via SCP or Hugging Face Hub. Use Axolotl or LLaMA-Factory config files to set hyperparameters, or write your own training script.
Train, evaluate, deploy
Launch training with real-time GPU monitoring. Evaluate against your test set, then push your model directly to a Runcrate serving instance.