unsloth_llama_4bit_lora_finetuning_on_alpaca_dataset.py

1from unsloth import FastLanguageModel
2import torch
3from trl import SFTTrainer
4from transformers import TrainingArguments
5from datasets import load_dataset
6
7# 1. Configuration
8max_seq_length = 2048
9dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
10load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
11
12# 2. Load model and tokenizer
13model, tokenizer = FastLanguageModel.from_pretrained(
14    model_name = "unsloth/llama-3-8b-bnb-4bit",
15    max_seq_length = max_seq_length,
16    dtype = dtype,
17    load_in_4bit = load_in_4bit,
18)
19
20# 3. Add LoRA weights
21model = FastLanguageModel.get_peft_model(
22    model,
23    r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
24    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
25                      "gate_proj", "up_proj", "down_proj",],
26    lora_alpha = 16,
27    lora_dropout = 0, # Supports any, but = 0 is optimized
28    bias = "none",    # Supports any, but = "none" is optimized
29    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
30    random_state = 3407,
31    use_rslora = False,
32    loftq_config = None,
33)
34
35# 4. Data Prep
36alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
37
38### Instruction:
39{}
40
41### Input:
42{}
43
44### Response:
45{}"""
46
47def formatting_prompts_func(examples):
48    instructions = examples["instruction"]
49    inputs       = examples["input"]
50    outputs      = examples["output"]
51    texts = []
52    for instruction, input, output in zip(instructions, inputs, outputs):
53        text = alpaca_prompt.format(instruction, input, output)
54        texts.append(text)
55    return { "text" : texts, }
56
57dataset = load_dataset("yahma/alpaca-cleaned", split = "train")
58dataset = dataset.map(formatting_prompts_func, batched = True,)
59
60# 5. Training
61trainer = SFTTrainer(
62    model = model,
63    tokenizer = tokenizer,
64    train_dataset = dataset,
65    dataset_text_field = "text",
66    max_seq_length = max_seq_length,
67    dataset_num_proc = 2,
68    packing = False, # Can make training 5x faster for short sequences.
69    args = TrainingArguments(
70        per_device_train_batch_size = 2,
71        gradient_accumulation_steps = 4,
72        warmup_steps = 5,
73        max_steps = 60,
74        learning_rate = 2e-4,
75        fp16 = not torch.cuda.is_bf16_supported(),
76        bf16 = torch.cuda.is_bf16_supported(),
77        logging_steps = 1,
78        optim = "adamw_8bit",
79        weight_decay = 0.01,
80        lr_scheduler_type = "linear",
81        seed = 3407,
82        output_dir = "outputs",
83    ),
84)
85
86trainer_stats = trainer.train()