startrail-org--pixelrag
wehub-resource-sync
53 行
2.3 KiB
Bash
可执行文件
53 行
2.3 KiB
Bash
可执行文件
#!/bin/bash
|
||
# Run the full LLM-judge matrix in parallel across 8 GPUs.
|
||
# 4 adapters (base, 2x_v2, 3x_v2, 4x_v2) × 4 k values (1,2,3,4) = 16 cells
|
||
# minus: base@k=1 (0.958), base@k=3 (0.892), 2x_v2@k=3 (0.930) already done = 13 new cells.
|
||
set -e
|
||
|
||
cd /scratch/users/zwcolin/cxr_embeds/cxr_embedding
|
||
source ~/.zshrc
|
||
|
||
DATA=/scratch/users/zwcolin/cxr_embeds/sft_data
|
||
OUT=/scratch/users/zwcolin/cxr_embeds/sft_output
|
||
LOGS=/scratch/users/zwcolin/cxr_embeds/cxr_embedding/sft/eval_out
|
||
|
||
launch() {
|
||
local gpu=$1 tag=$2 test_json=$3 adapter_arg=$4
|
||
CUDA_VISIBLE_DEVICES=$gpu uv run python sft/eval_multiimage.py \
|
||
--test-json "$test_json" $adapter_arg \
|
||
--tag "$tag" --n-examples 500 \
|
||
> "$LOGS/eval_${tag}.log" 2>&1 &
|
||
}
|
||
|
||
# ROUND 1: 8 cells
|
||
launch 0 "base_0x_k2" "$DATA/compressed_top6_0x/test_k2.json" ""
|
||
launch 1 "base_0x_k4" "$DATA/compressed_top6_0x/test_k4.json" ""
|
||
launch 2 "2x_v2_k1" "$DATA/compressed_top6_2x/test_k1.json" "--adapter $OUT/qwen3vl_top6_2x_v2"
|
||
launch 3 "2x_v2_k2" "$DATA/compressed_top6_2x/test_k2.json" "--adapter $OUT/qwen3vl_top6_2x_v2"
|
||
launch 4 "2x_v2_k4" "$DATA/compressed_top6_2x/test_k4.json" "--adapter $OUT/qwen3vl_top6_2x_v2"
|
||
launch 5 "3x_v2_k1" "$DATA/compressed_top6_3x/test_k1.json" "--adapter $OUT/qwen3vl_top6_3x_v2"
|
||
launch 6 "3x_v2_k2" "$DATA/compressed_top6_3x/test_k2.json" "--adapter $OUT/qwen3vl_top6_3x_v2"
|
||
launch 7 "3x_v2_k3" "$DATA/compressed_top6_3x/test.json" "--adapter $OUT/qwen3vl_top6_3x_v2"
|
||
wait
|
||
echo "=== ROUND 1 DONE ==="
|
||
|
||
# ROUND 2: 5 cells
|
||
launch 0 "3x_v2_k4" "$DATA/compressed_top6_3x/test_k4.json" "--adapter $OUT/qwen3vl_top6_3x_v2"
|
||
launch 1 "4x_v2_k1" "$DATA/compressed_top6_4x/test_k1.json" "--adapter $OUT/qwen3vl_top6_4x_v2"
|
||
launch 2 "4x_v2_k2" "$DATA/compressed_top6_4x/test_k2.json" "--adapter $OUT/qwen3vl_top6_4x_v2"
|
||
launch 3 "4x_v2_k3" "$DATA/compressed_top6_4x/test.json" "--adapter $OUT/qwen3vl_top6_4x_v2"
|
||
launch 4 "4x_v2_k4" "$DATA/compressed_top6_4x/test_k4.json" "--adapter $OUT/qwen3vl_top6_4x_v2"
|
||
wait
|
||
echo "=== ROUND 2 DONE ==="
|
||
|
||
# Summary
|
||
echo ""
|
||
echo "=== FINAL LLM-JUDGE ==="
|
||
for tag in base_0x_k2 base_0x_k4 \
|
||
2x_v2_k1 2x_v2_k2 2x_v2_k4 \
|
||
3x_v2_k1 3x_v2_k2 3x_v2_k3 3x_v2_k4 \
|
||
4x_v2_k1 4x_v2_k2 4x_v2_k3 4x_v2_k4; do
|
||
score=$(grep "llm_judge:" "$LOGS/eval_${tag}.log" 2>/dev/null | tail -1)
|
||
echo " $tag: $score"
|
||
done
|