From dc801c07cf38b0c495686463e6ca6f871a64440e Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Tue, 27 Jan 2026 09:57:37 -0600 Subject: Add collaborativeagents module and update gitignore - Add collaborativeagents subproject with adapters, agents, and evaluation modules - Update .gitignore to exclude large binary files (.whl, .tar), wandb logs, and results Co-Authored-By: Claude Opus 4.5 --- .../slurm/run_full_experiment_v2.sh | 61 ++++++++++++++++++++++ 1 file changed, 61 insertions(+) create mode 100755 collaborativeagents/slurm/run_full_experiment_v2.sh (limited to 'collaborativeagents/slurm/run_full_experiment_v2.sh') diff --git a/collaborativeagents/slurm/run_full_experiment_v2.sh b/collaborativeagents/slurm/run_full_experiment_v2.sh new file mode 100755 index 0000000..ab77478 --- /dev/null +++ b/collaborativeagents/slurm/run_full_experiment_v2.sh @@ -0,0 +1,61 @@ +#!/bin/bash +#SBATCH --job-name=full_exp_v2 +#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_v2_%j.out +#SBATCH --error=logs/full_exp_v2_%j.err + +# Full scale experiment v2 - with fixes: +# 1. Use stable datasets (math-500, humaneval) - avoid problematic ones +# 2. Reduced scale first (20 profiles, 15 sessions) to verify stability +# 3. Clear CUDA cache between methods + +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 v2 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}" + +# Disable peer-to-peer memory access to avoid nvlink errors +export CUDA_VISIBLE_DEVICES=0,1,2,3 +export NCCL_P2P_DISABLE=1 + +cd collaborativeagents/scripts + +# FULL SCALE v2: +# - 20 profiles (reduced from 30 for stability) +# - 15 sessions per profile (enough to show learning) +# - 2 stable datasets: math-500, humaneval +# - All 7 methods +echo "Running FULL SCALE v2 experiments..." +python run_experiments.py \ + --methods vanilla,all_memory,rag,rag_vector,contextual,reflection,reflection_grpo \ + --datasets math-500,humaneval \ + --n-profiles 20 \ + --n-sessions 15 \ + --profile-path ../data/complex_profiles_v2/profiles_100.jsonl \ + --output-dir ../results/full_experiment_v2_$(date +%Y%m%d_%H%M%S) + +echo "Full experiment v2 completed at $(date)" -- cgit v1.2.3