예제 소스

run_coord_r1.sh

#!/bin/bash
set -u
cd ~/project
PY=../go/.venv/bin/python
M=vlm_models/Qwen__Qwen3-VL-8B-Instruct
step() { echo "=== $(date '+%H:%M:%S') $*"; }
step base verify legacy;   $PY coord_omook.py verify --format legacy  --out logs/coord/base_legacy
step base verify aligned;  $PY coord_omook.py verify --format aligned --out logs/coord/base_aligned
step base real legacy;     $PY coord_omook.py realboard --format legacy  --n 150 --out logs/coord/real_base_legacy
step base real aligned;    $PY coord_omook.py realboard --format aligned --n 150 --out logs/coord/real_base_aligned
step dataset r1;           $PY coord_omook.py dataset --out-dir sft_data/coord_omook_r1 --n-samples 6000 --move-frac 0.2
step train r1
$PY train_lora_mlx.py --dataset-jsonl sft_data/coord_omook_r1/train.jsonl --val-jsonl sft_data/coord_omook_r1/val.jsonl \
  --model $M --output-dir vlm_models/lora_coord_omook_r1 --epochs 1 --batch-size 1 --grad-accum 8 --lr 5e-5 \
  --lora-rank 16 --lora-alpha 32 --max-seq-length 4096 --max-samples 3000 --grad-checkpoint
step r1 verify;            $PY coord_omook.py verify --format aligned --lora-dir vlm_models/lora_coord_omook_r1 --out logs/coord/r1_aligned
step r1 real;              $PY coord_omook.py realboard --format aligned --n 150 --lora-dir vlm_models/lora_coord_omook_r1 --out logs/coord/real_r1_aligned
step done