mirror of https://github.com/inclusionAI/AReaL
41 lines
905 B
Bash
41 lines
905 B
Bash
set -ex
|
|
|
|
|
|
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
|
|
export VLLM_LOGGING_LEVEL=DEBUG
|
|
|
|
MODEL_NAME_OR_PATH=$1
|
|
# OUTPUT_DIR=$1
|
|
MAX_GEN_TOKENS=${2:-4096}
|
|
DATA_NAME=${3:-"aime24,aime23"}
|
|
PROMPT_TYPE=${4:-"qwen-boxed"}
|
|
|
|
SPLIT="test"
|
|
NUM_TEST_SAMPLE=-1
|
|
OUTPUT_DIR=${5:-$MODEL_NAME_OR_PATH}
|
|
TASK=${6:-"math"}
|
|
|
|
# English open datasets
|
|
# DATA_NAME="math_500,math,gsm8k,train_amc_aime,aime24,amc23"
|
|
|
|
TOKENIZERS_PARALLELISM=false \
|
|
python3 -u ${TASK}_eval.py \
|
|
--model_name_or_path ${MODEL_NAME_OR_PATH} \
|
|
--data_name ${DATA_NAME} \
|
|
--output_dir ${OUTPUT_DIR} \
|
|
--split ${SPLIT} \
|
|
--prompt_type ${PROMPT_TYPE} \
|
|
--num_test_sample ${NUM_TEST_SAMPLE} \
|
|
--seed 0 \
|
|
--temperature 0 \
|
|
--n_sampling 1 \
|
|
--top_p 1 \
|
|
--start 0 \
|
|
--end -1 \
|
|
--use_vllm \
|
|
--max_tokens_per_call=$MAX_GEN_TOKENS \
|
|
--tensor_parallel_size 2 \
|
|
--save_outputs \
|
|
# --overwrite \
|
|
|