train_alizee_coder.py

# /// script
# dependencies = ["trl==0.11.4", "peft>=0.7.0", "trackio", "datasets", "transformers>=4.46.0", "accelerate", "bitsandbytes", "torch", "protobuf", "sentencepiece", "mistral-common>=1.5.0"]
# ///

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

print("="*50)
print("Starting Alizee Coder Devstral Training")
print("="*50)

# Configuration
MODEL_NAME = "mistralai/Devstral-Small-2505"
OUTPUT_REPO = "stmasson/alizee-coder-devstral-1-small"
DATASET_SIZE = 10000

# Verify HF_TOKEN
if not os.environ.get("HF_TOKEN"):
    raise ValueError("HF_TOKEN not set!")
print("HF_TOKEN verified")

print(f"Loading dataset nvidia/OpenCodeReasoning...")
try:
    dataset = load_dataset("nvidia/OpenCodeReasoning", "split_0", split="split_0")
    dataset = dataset.shuffle(seed=42).select(range(min(DATASET_SIZE, len(dataset))))
    print(f"Dataset loaded: {len(dataset)} examples")
except Exception as e:
    print(f"Error loading dataset: {e}")
    raise

# Split train/eval
dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
train_dataset = dataset_split["train"]
eval_dataset = dataset_split["test"]
print(f"Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")

# Format for code reasoning
def format_example(example):
    solution = example.get('solution', '') or ''
    output = example.get('output', '') or ''
    text = f"<s>[INST] Solve this programming problem with detailed reasoning:\n\n{example['input']}\n[/INST]\n\n**Reasoning:**\n{output}\n\n**Solution:**\n```python\n{solution}\n```</s>"
    return {"text": text}

print("Formatting dataset...")
train_dataset = train_dataset.map(format_example, remove_columns=train_dataset.column_names)
eval_dataset = eval_dataset.map(format_example, remove_columns=eval_dataset.column_names)
print("Dataset formatted")

# Load tokenizer
print(f"Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
print("Tokenizer loaded")

# 4-bit quantization
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_NAME}...")
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
print("Model loaded")

# LoRA configuration
lora_config = LoraConfig(
    r=32,
    lora_alpha=64,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.CAUSAL_LM,
)

# Training config
training_config = SFTConfig(
    output_dir="./alizee-coder-devstral-1-small",
    num_train_epochs=1,
    per_device_train_batch_size=1,
    per_device_eval_batch_size=1,
    gradient_accumulation_steps=16,
    gradient_checkpointing=True,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    max_seq_length=4096,
    logging_steps=10,
    save_strategy="steps",
    save_steps=200,
    eval_strategy="steps",
    eval_steps=200,
    bf16=True,
    push_to_hub=True,
    hub_model_id=OUTPUT_REPO,
    hub_strategy="every_save",
    report_to="trackio",
    run_name="alizee-coder-devstral-1-small",
)

print("Initializing trainer...")
trainer = SFTTrainer(
    model=model,
    args=training_config,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    peft_config=lora_config,
    tokenizer=tokenizer,
    dataset_text_field="text",
)

print("="*50)
print("STARTING TRAINING")
print("="*50)
trainer.train()

print("Pushing to Hub...")
trainer.push_to_hub()
print(f"Done! Model: https://huggingface.co/{OUTPUT_REPO}")