blob: bbf53fc5704c86d4ecee2b504fdb28da263842bf (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
|
#!/bin/bash
#SBATCH --job-name=run_collab_baselines
#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/run_collab_baselines_%j.out
#SBATCH --error=logs/run_collab_baselines_%j.err
# Run CollaborativeAgents baselines on 4x A100 80GB
# - contextual: Full history in context (summarize on overflow)
# - reflection: CollaborativeAgents' agent_notes approach
# - reflection_grpo: Reflection + GRPO training (with_proper_scaffolding)
set -e
cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model
# Create logs and results directories
mkdir -p collaborativeagents/slurm/logs
mkdir -p collaborativeagents/results
echo "Starting CollaborativeAgents baselines at $(date)"
echo "Job ID: $SLURM_JOB_ID"
echo "Node: $SLURMD_NODENAME"
echo "GPUs: $CUDA_VISIBLE_DEVICES"
# Activate environment
source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
conda activate eval
# Check GPU availability
nvidia-smi
# Redirect HF cache to project space (avoid home quota issues)
export HF_HOME=/projects/bfqt/users/yurenh2/.cache/huggingface
mkdir -p $HF_HOME
# Add project to path
export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"
# Run experiments
cd collaborativeagents/scripts
# Run the 3 CollaborativeAgents baselines
echo "Running contextual, reflection, reflection_grpo baselines..."
python run_experiments.py \
--methods contextual,reflection,reflection_grpo \
--datasets math-500 \
--n-profiles 20 \
--n-sessions 5 \
--profile-path ../data/complex_profiles_v2/profiles_100.jsonl \
--output-dir ../results/collab_baselines_$(date +%Y%m%d_%H%M%S)
echo "CollaborativeAgents baselines completed at $(date)"
|
