Fine-Tuning Llama 4 on Wollnut Labs: A Step-by-Step Guide

Prerequisites

Step 1: Launch a GPU Instance

Your instance will be ready in about 60 seconds.

Step 2: Connect via SSH

ssh -i ~/.ssh/your_key root@YOUR_INSTANCE_IP

Step 3: Install Fine-Tuning Dependencies

pip install trl peft datasets bitsandbytes accelerate

Step 4: Upload Your Dataset

scp -i ~/.ssh/your_key dataset.jsonl root@YOUR_IP:/workspace/

Step 5: Run Fine-Tuning

Create a training script:

from trl import SFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig
from datasets import load_dataset

model_name = "meta-llama/Llama-4-Scout-17B"
dataset = load_dataset("json", data_files="dataset.jsonl")

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
)

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    peft_config=lora_config,
    max_seq_length=2048,
    tokenizer=tokenizer,
)

trainer.train()
trainer.save_model("./llama-finetuned")

Step 6: Export Your Model

scp -r -i ~/.ssh/your_key root@YOUR_IP:/workspace/llama-finetuned ./

Step 7: Stop Your Instance

Don't forget to stop your instance when training is complete! Billing stops immediately.

Cost Estimate

Fine-tuning Llama 4 Scout 17B on a typical dataset (10k examples, 3 epochs) takes about 2-4 hours on an H100. At $2.25/hr, that's roughly $4.50–$9.00 total.