Llm finetuning

LLM Fine-Tuning

Overview

Fine-tuning adapts a base LLM to a task or style. For almost all practitioners the right default is parameter-efficient fine-tuning (LoRA/QLoRA) — it trains <1% of weights, fits on a single GPU, and avoids catastrophic forgetting. Reach for full fine-tuning only with strong justification and budget.

When to use

• A base/instruct model is close but needs domain style, format, or behavior.
• You have a few hundred to tens of thousands of quality examples.

First decide: do you even need to fine-tune?

Try prompting + few-shot + RAG first (see the rag-pipeline skill). Fine-tune when you need consistent format/style, lower latency/cost than long prompts, or to internalize a large example set.

Method selection

Method	Trains	VRAM	Use when
Full FT	100%	Very high	Large data + budget, max quality
LoRA	~0.1-1%	Moderate	Default for most tasks
QLoRA	~0.1-1%	Low (4-bit)	Single consumer GPU

Dataset formatting (chat / instruction)

Use the model's chat template; quality > quantity. A few thousand clean, diverse examples beat a noisy 100k dump.

def to_chat(example):
    return {"messages": [
        {"role": "system", "content": "You are a helpful support agent."},
        {"role": "user", "content": example["question"]},
        {"role": "assistant", "content": example["answer"]},
    ]}

QLoRA with TRL

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import torch

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")

peft_cfg = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none",
                      task_type="CAUSAL_LM",
                      target_modules=["q_proj", "k_proj", "v_proj", "o_proj"])

trainer = SFTTrainer(
    model=model, train_dataset=train_ds, eval_dataset=eval_ds, peft_config=peft_cfg,
    args=SFTConfig(per_device_train_batch_size=2, gradient_accumulation_steps=8,
                   learning_rate=2e-4, num_train_epochs=2, warmup_ratio=0.03,
                   lr_scheduler_type="cosine", bf16=True, logging_steps=10,
                   eval_strategy="steps", eval_steps=100, output_dir="out"),
)
trainer.train()

Key hyperparameters

• LR: 1e-4 to 3e-4 for LoRA (higher than full FT).
• Epochs: 1-3 — more overfits and degrades general ability.
• LoRA rank r: 8-32 typical; raise for harder tasks.
• Effective batch: use gradient accumulation to reach 16-32.

Evaluating a fine-tune

• Hold out a real test set; track eval loss for overfitting.
• Judge on task metrics (exact match, rubric/LLM-judge), not just loss.
• Check for regression on general benchmarks — fine-tuning can break unrelated abilities.

Pitfalls

• Too many epochs / LR too high → overfit, repetitive or degenerate outputs.
• Wrong chat template → the model learns malformed formatting.
• Tiny/dirty dataset → memorization, not generalization.
• Skipping a base-model baseline → you can't prove the fine-tune helped.

Hand-off

Merged or adapter weights + eval report, logged via experiment-tracking and deployed via model-serving.