#!/bin/bash
#SBATCH --job-name=sft_refl
#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=logs/sft_reflection_%j.out
#SBATCH --error=logs/sft_reflection_%j.err

set -e
cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model
mkdir -p collaborativeagents/slurm/logs collaborativeagents/training/outputs collaborativeagents/training/training_data

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:${PYTHONPATH}"
export WANDB_PROJECT="collaborative-agent-reflection-sft"

# Step 1: Generate training data from completed experiments
echo "=== Step 1: Generating training data ==="
python collaborativeagents/training/generate_training_data.py \
    --results-dir collaborativeagents/results \
    --output-dir collaborativeagents/training/training_data

# Step 2: Run SFT training using TRL
echo "=== Step 2: Running SFT training ==="
python collaborativeagents/training/train_sft.py \
    --model-path /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct \
    --data-path collaborativeagents/training/training_data/sft_training_data.json \
    --output-dir collaborativeagents/training/outputs/sft_reflection \
    --num-epochs 4 \
    --learning-rate 1e-6 \
    --batch-size 1 \
    --gradient-accumulation 64

echo "=== SFT Training Complete ==="
echo "Model saved to: collaborativeagents/training/outputs/sft_reflection"