train_mistral.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "torch>=2.0.0", "transformers>=4.40.0", "accelerate>=0.20.0", "bitsandbytes>=0.41.0", "protobuf>=3.20.0", "sentencepiece>=0.1.99"]
# ///

"""Fine-tune Mistral-7B-Instruct-v0.3 on NATO doctrine dataset."""

import os
from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from huggingface_hub import login
import torch
import trackio

# Authenticate with HF Hub using token from secrets
hf_token = os.environ.get("HF_TOKEN")
if hf_token:
    login(token=hf_token)
    print("✓ Logged in to Hugging Face Hub")
else:
    print("⚠ Warning: HF_TOKEN not found in environment")

# Model ID
model_id = "mistralai/Mistral-7B-Instruct-v0.3"

# Load tokenizer
print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

# Load model with 4-bit quantization
print("Loading model with 4-bit quantization...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
model.config.use_cache = False
model.gradient_checkpointing_enable()

print(f"✓ Model loaded: {model_id}")

# Load dataset from HF Hub
print("\nLoading NATO doctrine dataset...")
dataset = load_dataset("AndreasThinks/nato-doctrine-sft", split="train")
dataset_test = load_dataset("AndreasThinks/nato-doctrine-sft", split="test")

print(f"✓ Train set: {len(dataset)} examples")
print(f"✓ Test set: {len(dataset_test)} examples")

# Configure LoRA for efficient fine-tuning
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 configuration
training_args = SFTConfig(
    output_dir="nato-ministral-3b",

    # Model saving
    push_to_hub=True,
    hub_model_id="AndreasThinks/mistral-7b-nato-doctrine",
    hub_strategy="every_save",
    hub_private_repo=False,

    # Training parameters
    num_train_epochs=3,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=8,  # Effective batch size = 16
    gradient_checkpointing=True,

    # Learning rate
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,

    # Optimization
    optim="adamw_torch",
    weight_decay=0.01,
    max_grad_norm=1.0,

    # Evaluation
    eval_strategy="steps",
    eval_steps=50,

    # Logging and saving
    logging_steps=10,
    save_strategy="steps",
    save_steps=100,
    save_total_limit=3,

    # Monitoring with Trackio
    report_to="trackio",
    run_name="nato-mistral-7b-v1",
    project="nato-doctrine-training",

    # Other
    bf16=True,  # Use bfloat16 for better stability
    seed=42,
)

# Initialize trainer with loaded model
print("\n✓ Initializing SFT trainer...")
trainer = SFTTrainer(
    model=model,
    processing_class=tokenizer,
    train_dataset=dataset,
    eval_dataset=dataset_test,
    peft_config=peft_config,
    args=training_args,
)

# Start training
print("\n✓ Starting training...")
print(f"  Model: mistralai/Mistral-7B-Instruct-v0.3")
print(f"  Training examples: {len(dataset)}")
print(f"  Test examples: {len(dataset_test)}")
print(f"  Epochs: 3")
print(f"  LoRA rank: 16")
print(f"  Output: AndreasThinks/mistral-7b-nato-doctrine\n")

trainer.train()

# Save final model
print("\n✓ Training complete! Saving final model...")
trainer.push_to_hub()

print("\n✅ Fine-tuning complete!")
print(f"  Model: https://huggingface.co/AndreasThinks/mistral-7b-nato-doctrine")
print(f"  Base: mistralai/Mistral-7B-Instruct-v0.3")
print(f"  Trackio: Check your dashboard for metrics")