unsloth-mcp-server: Skill for Claude Code

Training Optimization

Master advanced techniques for efficient, high-quality LLM fine-tuning.

Overview

Fine-tuning is an art. Optimize:

• Learning rates - Schedulers, warmup, optimal values
• LoRA configuration - Rank, alpha, target modules
• Batch optimization - Size, accumulation, sequence length
• Precision - FP16, BF16, mixed precision
• Gradient strategies - Checkpointing, clipping, accumulation
• Hyperparameter tuning - Grid search, Bayesian optimization
• Monitoring - WandB, TensorBoard, loss curves
• Quality - Prevent overfitting, improve convergence

Quick Start

Optimal Default Configuration

from unsloth import FastLanguageModel
from trl import SFTTrainer
from transformers import TrainingArguments

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/Llama-3.2-7B-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True,
    use_gradient_checkpointing="unsloth"  # Memory efficient
)

# Configure LoRA (optimal defaults)
model = FastLanguageModel.get_peft_model(
    model,
    r=16,                    # LoRA rank (8-64)
    lora_alpha=16,          # Alpha = rank typically works well
    lora_dropout=0,         # 0 for Unsloth (already efficient)
    target_modules=[        # All attention + MLP for best quality
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    bias="none",
    use_gradient_checkpointing="unsloth"
)

# Training arguments (optimal defaults)
training_args = TrainingArguments(
    output_dir="./outputs",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,        # Effective batch = 2*4 = 8
    num_train_epochs=3,
    learning_rate=2e-4,                   # 2e-4 is a sweet spot
    lr_scheduler_type="cosine",           # Cosine decay
    warmup_ratio=0.03,                    # 3% warmup
    fp16=not torch.cuda.is_bf16_supported(),
    bf16=torch.cuda.is_bf16_supported(),  # BF16 if available
    logging_steps=10,
    optim="adamw_8bit",                   # 8-bit AdamW (memory efficient)
    save_strategy="epoch",
    save_total_limit=3
)

# Train
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    max_seq_length=2048,
    args=training_args
)

trainer.train()