"""
PhD Research OS — SFT Training Script
=======================================
Fine-tunes Qwen2.5-3B-Instruct using QLoRA on multi-task scientific research data.

Tasks trained:
  1. Scientific Claim Extraction (structured JSON output)
  2. Epistemic Classification (Fact/Interpretation/Hypothesis/Conflict_Hypothesis)
  3. Confidence Scoring (evidence_strength × study_quality × journal_tier × completeness)
  4. Contradiction Detection (claim pair → conflict analysis)
  5. Query Decomposition (broad question → sub-queries)
  6. Decision Object Generation (knowledge gaps → proposed research actions)

Base model: Qwen/Qwen2.5-3B-Instruct
Method: QLoRA (r=64, all-linear) following "LoRA Without Regret" recipe
Reference: arxiv:2212.05238 (LLM-NERRE), arxiv:2401.00579 (multi-task biomedical SFT)

Usage:
  pip install torch transformers trl peft datasets bitsandbytes accelerate trackio
  python train.py
"""

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

MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"
DATASET_NAME = "nkshirsa/phd-research-os-sft-data"
OUTPUT_DIR = "./phd-research-os-brain"
HUB_MODEL_ID = "nkshirsa/phd-research-os-brain"

trackio.init(project="phd-research-os-training", run="sft-qwen25-3b-qlora-v1")

print("Loading dataset...")
dataset = load_dataset(DATASET_NAME)
train_dataset = dataset["train"]
eval_dataset = dataset["test"]
print(f"Train: {len(train_dataset)} examples, Eval: {len(eval_dataset)} examples")

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

peft_config = LoraConfig(
    r=64, lora_alpha=16, lora_dropout=0.05, bias="none",
    task_type="CAUSAL_LM", target_modules="all-linear")

training_args = SFTConfig(
    output_dir=OUTPUT_DIR,
    num_train_epochs=3,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.05,
    weight_decay=0.01,
    max_grad_norm=1.0,
    bf16=True,
    gradient_checkpointing=True,
    max_length=2048,
    model_init_kwargs={"quantization_config": bnb_config, "torch_dtype": torch.bfloat16},
    assistant_only_loss=True,
    logging_steps=5, logging_first_step=True, disable_tqdm=True,
    report_to=["tensorboard"], logging_dir=f"{OUTPUT_DIR}/logs",
    eval_strategy="steps", eval_steps=50,
    save_strategy="steps", save_steps=100, save_total_limit=3,
    load_best_model_at_end=True, metric_for_best_model="eval_loss", greater_is_better=False,
    push_to_hub=True, hub_model_id=HUB_MODEL_ID, hub_strategy="every_save",
    seed=42, data_seed=42)

trainer = SFTTrainer(
    model=MODEL_NAME, args=training_args,
    train_dataset=train_dataset, eval_dataset=eval_dataset,
    peft_config=peft_config)

trainable = sum(p.numel() for p in trainer.model.parameters() if p.requires_grad)
total = sum(p.numel() for p in trainer.model.parameters())
print(f"Model: {MODEL_NAME} | Total: {total:,} | Trainable: {trainable:,} ({100*trainable/total:.2f}%)")

train_result = trainer.train()
trainer.save_model()
trainer.push_to_hub(commit_message="Final model: PhD Research OS Brain v1")

print(f"\nTraining complete! Model at: https://huggingface.co/{HUB_MODEL_ID}")
for k, v in train_result.metrics.items():
    print(f"  {k}: {v}")