# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.36.0",
#     "accelerate>=0.24.0",
#     "trackio",
#     "datasets",
# ]
# ///

import os
os.environ["HF_HUB_DISABLE_XET"] = "1"

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

print("Loading memory-agent-sft-v1 dataset...")
dataset = load_dataset("erik1988/memory-agent-sft-v1", split="train")
print(f"Dataset loaded: {len(dataset)} examples")

print(f"Train: {len(dataset)} examples (no eval split to save GPU memory)")

config = SFTConfig(
    output_dir="elias-memory-agent-v2",
    push_to_hub=True,
    hub_model_id="erik1988/elias-memory-agent-v2",
    hub_strategy="every_save",
    num_train_epochs=3,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    learning_rate=2e-5,
    fp16=True,
    max_length=512,
    logging_steps=5,
    save_strategy="steps",
    save_steps=50,
    save_total_limit=2,
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    gradient_checkpointing=True,
    report_to="trackio",
    project="elias-identity",
    run_name="memory-agent-sft-v2-retry",
)

peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"],
)

print("Initializing trainer with Qwen2.5-3B...")
trainer = SFTTrainer(
    model="Qwen/Qwen2.5-3B",
    train_dataset=dataset,
    args=config,
    peft_config=peft_config,
)

print("Starting training...")
trainer.train()
print("Pushing to Hub...")
trainer.push_to_hub()
trackio.finish()
print("DONE: Memory Agent v2")