r18 scripts sync
Browse files- r18_materials/r17_train.sh +18 -0
r18_materials/r17_train.sh
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# r17_train.sh — r17 深思熟虑强化训练 (由 rc-bg 后台 job 启动, exec 保证进程存活)
|
| 3 |
+
LOG=/mnt/workspace/r17_train.log
|
| 4 |
+
{
|
| 5 |
+
echo "[$(date '+%F %T')] ===== r17 train wrapper start ====="
|
| 6 |
+
echo "cwd=$(pwd)"; echo "whoami=$(whoami)"; echo "python3=$(which python3)"
|
| 7 |
+
echo "GPU: $(nvidia-smi --query-gpu=memory.used,memory.total --format=csv,noheader)"
|
| 8 |
+
} >> $LOG 2>&1
|
| 9 |
+
cd /mnt/workspace/pnet_repo/pondernet || { echo "[$(date '+%F %T')] CD_FAIL" >> $LOG; exit 1; }
|
| 10 |
+
echo "[$(date '+%F %T')] entering train_ponder_moe.py" >> $LOG
|
| 11 |
+
echo $$ > /mnt/workspace/r17_train.pid
|
| 12 |
+
PYTHONUNBUFFERED=1 exec python3 train_ponder_moe.py \
|
| 13 |
+
--model out_2b/pnet2b_moe --init-from out_2b/ckpt_2b_r16 \
|
| 14 |
+
--data out_2b/data_agent/mix_r17.jsonl --out out_2b/ckpt_2b_r17 \
|
| 15 |
+
--lora 16 --lr 1.5e-4 --lr-lora 4e-5 --epochs 2 --accum 4 \
|
| 16 |
+
--difficulty-prior easy:0.85,medium:0.35,hard:0.1 --prior-p 0.5 \
|
| 17 |
+
--ponder-loss-beta 0.05 --step-ce-weight 0.5 --grad-ckpt \
|
| 18 |
+
--max-len 1280 --save-every 150 >> $LOG 2>&1
|