train_aviation.py

# /// script
# dependencies = [
#     "torch", 
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.46.0",
#     "huggingface_hub>=0.26.0",
#     "accelerate>=0.24.0",
#     "trackio",
#     "bitsandbytes",
#     "scipy",
# ]
# ///

import trackio
import torch
from datasets import load_dataset
from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model
from trl import SFTTrainer, SFTConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# Model ID
model_id = "mistralai/Mistral-3-14B-Reasoning-2512"

# Load dataset
print("📦 Loading dataset...")
dataset = load_dataset("sakharamg/AviationQA", split="train")

# Limit dataset size for reasonable training time (e.g., 10k examples)
# 1M rows is too large for a single generic fine-tuning job without massive compute.
print("✂️ Subsampling dataset to 10,000 examples for efficiency...")
dataset = dataset.shuffle(seed=42).select(range(10000))

# Map to chat format
print("🔄 Mapping dataset...")
def to_messages(example):
    return {
        "messages": [
            {"role": "user", "content": example["Question"]},
            {"role": "assistant", "content": example["Answer"]}
        ]
    }
dataset = dataset.map(to_messages, remove_columns=dataset.column_names)

# Split
print("🔀 Creating train/eval split...")
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]

# Quantization Config (4-bit for memory efficiency)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

# Load Model
print(f"🤖 Loading model {model_id}...")
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="eager" # Default attention for compatibility
)
model = prepare_model_for_kbit_training(model)

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token = tokenizer.eos_token
# Fix for some models that miss chat_template or padding
if tokenizer.chat_template is None:
    tokenizer.chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"

# LoRA Config
peft_config = 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", "gate_proj", "up_proj", "down_proj"],
)

# Training Config
config = SFTConfig(
    output_dir="Mistral-3-14B-AviationQA-SFT",
    push_to_hub=True,
    hub_model_id="sunkencity/Mistral-3-14B-AviationQA-SFT",
    hub_strategy="every_save",
    num_train_epochs=1,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    fp16=False,
    bf16=True,
    logging_steps=10,
    save_strategy="steps",
    save_steps=100,
    eval_strategy="steps",
    eval_steps=100,
    report_to="trackio",
    project="aviation-qa-tuning",
    run_name="mistral-14b-sft-v1",
    max_seq_length=2048,
    dataset_kwargs={"add_special_tokens": False} # Let tokenizer handle chat template
)

# Trainer
trainer = SFTTrainer(
    model=model,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    args=config,
    peft_config=peft_config,
    tokenizer=tokenizer,
)

print("🚀 Starting training...")
trainer.train()

print("💾 Pushing to Hub...")
trainer.push_to_hub()