test_training_llama_small_batch.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "transformers>=4.40.0", "datasets>=2.18.0", "accelerate>=0.28.0"]
# ///

from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import trackio

print("=" * 80)
print("TEST RUN: Biomedical Llama Fine-Tuning (100 examples)")
print("=" * 80)

print("\n[1/4] Loading dataset...")
dataset = load_dataset("panikos/biomedical-llama-training")

# Use first 100 examples for test
train_dataset = dataset["train"].select(range(100))
eval_dataset = dataset["validation"].select(range(20))

print(f"  Train: {len(train_dataset)} examples")
print(f"  Eval: {len(eval_dataset)} examples")

print("\n[2/4] Configuring LoRA...")
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)
print("  LoRA rank: 16, alpha: 32")

print("\n[3/4] Initializing trainer...")
trainer = SFTTrainer(
    model="meta-llama/Llama-3.1-8B-Instruct",
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    peft_config=lora_config,
    args=SFTConfig(
        output_dir="llama-biomedical-test",
        num_train_epochs=1,
        per_device_train_batch_size=1,  # REDUCED from 2 to 1
        gradient_accumulation_steps=8,  # INCREASED from 4 to 8
        learning_rate=2e-4,
        lr_scheduler_type="cosine",
        warmup_ratio=0.1,
        logging_steps=5,
        eval_strategy="steps",
        eval_steps=20,
        save_strategy="epoch",
        push_to_hub=True,
        hub_model_id="panikos/llama-biomedical-test",
        hub_private_repo=True,
        bf16=True,
        gradient_checkpointing=False,  # DISABLED for stability
        report_to="trackio",
        project="biomedical-llama-training",
        run_name="test-run-100-examples-v3"
    )
)

print("\n[4/4] Starting training...")
print("  Model: meta-llama/Llama-3.1-8B-Instruct")
print("  Method: SFT with LoRA")
print("  Epochs: 1")
print("  Batch size: 1 x 8 = 8 (effective) - optimized for memory")
print("  Gradient checkpointing: DISABLED")
print()

trainer.train()

print("\n" + "=" * 80)
print("Pushing model to Hub...")
print("=" * 80)
trainer.push_to_hub()

print("\n" + "=" * 80)
print("TEST COMPLETE!")
print("=" * 80)
print("\nModel: https://huggingface.co/panikos/llama-biomedical-test")
print("Dashboard: https://panikos-trackio.hf.space/")
print("Ready for full production training!")