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

"""
Agent Zero SFT v2: LiquidAI/LFM2.5-1.2B-Instruct
LoRA fine-tuning on mixed agent-zero-sft-v2 dataset.

Changes from v1:
  - Mixed dataset: 40% agent + 40% math (MetaMathQA) + 20% general (OpenHermes)
  - LoRA r=8 (was 16), alpha=16 (was 32) — reduced rank to prevent overfitting
  - 2 epochs (was 3)
  - lr=1e-4 (was 2e-4) — gentler updates
"""

import os

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

token = os.getenv("HF_TOKEN")
if token:
    login(token=token)

# Load v2 mixed dataset
print("Loading dataset...")
train_ds = load_dataset("wheattoast11/agent-zero-sft-v2", split="train")
val_ds = load_dataset("wheattoast11/agent-zero-sft-v2", split="validation")
print(f"Train: {len(train_ds)}, Val: {len(val_ds)}")

config = SFTConfig(
    output_dir="agent-zero-lfm-1.2b-v2",
    push_to_hub=True,
    hub_model_id="wheattoast11/agent-zero-lfm-1.2b-v2",
    hub_strategy="every_save",
    hub_private_repo=True,

    # v2: 2 epochs (was 3)
    num_train_epochs=2,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,  # effective batch size 16
    # v2: lr=1e-4 (was 2e-4)
    learning_rate=1e-4,
    bf16=True,

    logging_steps=10,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,

    eval_strategy="steps",
    eval_steps=200,

    warmup_ratio=0.1,
    lr_scheduler_type="cosine",

    report_to="trackio",
    project="agent-zero-finetune",
    run_name="lfm-1.2b-sft-v2",
)

# v2: r=8 (was 16), alpha=16 (was 32), ratio stays 2.0
peft_config = LoraConfig(
    r=8,
    lora_alpha=16,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
)

print("Initializing trainer...")
trainer = SFTTrainer(
    model="LiquidAI/LFM2.5-1.2B-Instruct",
    train_dataset=train_ds,
    eval_dataset=val_ds,
    args=config,
    peft_config=peft_config,
)

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

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

trackio.finish()
print("Done! Model at: https://huggingface.co/wheattoast11/agent-zero-lfm-1.2b-v2")