File size: 1,502 Bytes
c26b74d | 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 | #!/bin/bash
# stage5: slate models from the train-only base m_lr8
# c28: 28 layers, 60% of val/test -> hotkey 1/1 (uid 169)
# b16: 16 layers, 35% of val/test -> hotkey hata/4 (uid 52)
# d8 : 8 layers, 70% of val/test -> hotkey hata/5 (uid 53)
set -e
cd $HOME/mt92
export TOKENIZERS_PARALLELISM=false
set -a; . ./.env; set +a
curl -s -L https://huggingface.co/dasLOL/mt-extract-3g/resolve/main/manifest.json -o manifest_template.json
[ -d p16m ] || ./venv/bin/python prune.py m_lr8 p16m 16 2>&1 | grep -E 'dropping|saved'
[ -d p8m ] || ./venv/bin/python prune.py m_lr8 p8m 8 2>&1 | grep -E 'dropping|saved'
train() { # name base epochs set
echo "##### $1: base $2, $3 epochs, $4"
[ -d $1 ] || OUT=$1 BASE=$2 EPOCHS=$3 LR=8e-5 TRAIN=$4 ./venv/bin/python sft.py 2>&1 | tr '\r' '\n' | grep -E "train_loss|saved|Error|error"
for s in val_all test_all gen_holdout; do ./venv/bin/python evalset.py $1 ex/$s.jsonl; done
./venv/bin/python backup_now.py $1 > backup_$1.log 2>&1 && echo "$1 backed up"
bash ./build_final.sh $1 32 > build_$1.log 2>&1 && grep -E "MICRO|artifact|TTFT" build_$1.log
cp manifest_template.json onnx_$1/final/manifest.json
./venv/bin/python calib.py onnx_$1/final 60 > calib_$1.log 2>&1; grep -v "^probe" calib_$1.log | tr -d "\n"; echo
./venv/bin/python up_art.py onnx_$1/final onnx_$1 > up_$1.log 2>&1 && echo "staged onnx_$1"
}
train c28 m_lr8 4 ex/full60.jsonl
train b16 p16m 4 ex/full35.jsonl
train d8 p8m 8 ex/full70.jsonl
echo STAGE5_DONE
|