#!/bin/bash
#SBATCH --job-name=sft_train
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:4
#SBATCH --mem=200G
#SBATCH --time=24:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/sft_train_%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/sft_train_%j.err

# SFT Training only (data already exists)

set -e
cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model

source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PWD}/collaborativeagents/scripts:${PYTHONPATH}"
export WANDB_PROJECT="collaborative-agent-reflection-sft"

echo "=== SFT Training ==="
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv

DATA_PATH="collaborativeagents/training/training_data/sft_training_data.json"
MODEL_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
OUTPUT_DIR="collaborativeagents/training/outputs/sft_reflection"

echo "Data: $DATA_PATH"
echo "Model: $MODEL_PATH"
echo "Output: $OUTPUT_DIR"

# Check data exists
if [ ! -f "$DATA_PATH" ]; then
    echo "ERROR: Training data not found at $DATA_PATH"
    exit 1
fi

echo ""
echo "Training data size: $(wc -c < $DATA_PATH) bytes"
python -c "import json; d=json.load(open('$DATA_PATH')); print(f'Training examples: {len(d)}')"

mkdir -p $OUTPUT_DIR

echo ""
echo "Starting SFT training..."
python collaborativeagents/training/train_sft.py \
    --model-path $MODEL_PATH \
    --data-path $DATA_PATH \
    --output-dir $OUTPUT_DIR \
    --num-epochs 4 \
    --learning-rate 1e-6 \
    --batch-size 1 \
    --gradient-accumulation 64

echo ""
echo "=== SFT Training Complete ==="
echo "Model saved to: $OUTPUT_DIR"
date