train_aviation.py

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

import trackio
import torch
import os
from huggingface_hub import list_repo_files

model_id = "mistralai/Ministral-3-14B-Reasoning-2512" 

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, 
    AutoConfig, 
    AutoModel,
    MistralConfig 
)

# ------------------------------------------------------------------
# CRITICAL FIX: Manually wire the Ministral3 Hierarchy
# ------------------------------------------------------------------
print("🔧 Starting Manual Registration/Wiring...")

try:
    # --- 1. Inner Text Model (Ministral) ---
    from transformers.models.ministral.configuration_ministral import MinistralConfig
    from transformers.models.ministral.modeling_ministral import MinistralModel
    
    # Compatibility Config for Inner Model
    class Ministral3CompatConfig(MinistralConfig):
        model_type = "ministral3" 
        def __init__(self, **kwargs):
            super().__init__(**kwargs)
            if not hasattr(self, 'sliding_window') or self.sliding_window is None:
                self.sliding_window = 4096 
            if not hasattr(self, 'layer_types'):
                self.layer_types = ["sliding_attention"] * getattr(self, "num_hidden_layers", 40)

    # Compatibility Model for Inner Model
    class Ministral3CompatModel(MinistralModel):
        config_class = Ministral3CompatConfig

    # Register Inner Components
    AutoConfig.register("ministral3", Ministral3CompatConfig)
    AutoModel.register(Ministral3CompatConfig, Ministral3CompatModel)
    print("   ✅ Registered Inner: 'ministral3' -> Ministral3CompatModel")


    # --- 2. Outer Multimodal Model (Mistral3) ---
    from transformers.models.mistral3.configuration_mistral3 import Mistral3Config
    from transformers.models.mistral3.modeling_mistral3 import Mistral3ForConditionalGeneration

    # Register Outer Components with AutoModelForCausalLM
    AutoModelForCausalLM.register(Mistral3Config, Mistral3ForConditionalGeneration)
    print("   ✅ Registered Outer: Mistral3Config -> Mistral3ForConditionalGeneration")

except ImportError as e:
    print(f"   ❌ Failed to import/register classes: {e}")
    print("   ⚠️ This usually means the transformers version is too old or incompatible.")

# ------------------------------------------------------------------
# Standard Training Setup
# ------------------------------------------------------------------

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

print("✂️ Subsampling dataset to 10,000 examples for efficiency...")
dataset = dataset.shuffle(seed=42).select(range(12000)) 

print("🧹 Filtering invalid examples...")
dataset = dataset.filter(lambda x: x["Question"] and x["Answer"] and len(x["Question"].strip()) > 0 and len(x["Answer"].strip()) > 0)
if len(dataset) > 10000:
    dataset = dataset.select(range(10000))

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)

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"]

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

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" 
)
model = prepare_model_for_kbit_training(model)

tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token = tokenizer.eos_token
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 %}"

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"],
)

config = SFTConfig(
    output_dir="Ministral-3-14B-AviationQA-SFT",
    push_to_hub=True,
    hub_model_id="sunkencity/Ministral-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_length=2048,
    dataset_kwargs={"add_special_tokens": False} 
)

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

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

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