#!/bin/bash
#SBATCH --job-name=full_exp
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA100x4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:nvidia_a100:4
#SBATCH --mem=200G
#SBATCH --time=48:00:00
#SBATCH --output=logs/full_exp_%j.out
#SBATCH --error=logs/full_exp_%j.err

# Full scale experiment with 70B user model
# ORIGINAL CONFIG - DO NOT CHANGE

set -e

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

mkdir -p collaborativeagents/slurm/logs
mkdir -p collaborativeagents/results

echo "Starting FULL SCALE experiment at $(date)"
echo "Job ID: $SLURM_JOB_ID"
echo "Node: $SLURMD_NODENAME"
echo "GPUs: $CUDA_VISIBLE_DEVICES"

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

nvidia-smi

export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
mkdir -p $HF_HOME

export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"

# Fix for nvlink errors
export NCCL_P2P_DISABLE=1

cd collaborativeagents/scripts

# ORIGINAL FULL SCALE CONFIG:
# - 30 profiles
# - 20 sessions per profile
# - 4 challenging datasets: gpqa, aime, math-hard, humaneval
# - All 7 methods
echo "Running FULL SCALE experiments with 70B user model..."
python run_experiments.py \
    --methods vanilla,all_memory,rag,rag_vector,contextual,reflection,reflection_grpo \
    --datasets mmlu,aime,math-hard,humaneval \
    --n-profiles 30 \
    --n-sessions 20 \
    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl \
    --output-dir ../results/full_experiment_$(date +%Y%m%d_%H%M%S)

echo "Full experiment completed at $(date)"