evaluate_online.sh

export HF_HUB_OFFLINE=1
export DATASET_ROOT="./evaluate/dataset/C_C++_Java_Python"
export RESULT_ROOT="./evaluate/result"
export EVAL_SCRIPT="./dataset_eval.py"

GPT-3.5-turbo
export KEY=""
export URL="https://api.chatanywhere.tech/v1/"
export MODEL="gpt-3.5-turbo"

mkdir -p $RESULT_ROOT/$MODEL
export INCLUDE_MSG=no
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/nomsg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/nomsg.log
export INCLUDE_MSG=yes
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/msg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/msg.log

# DeepSeek-V3
export KEY=""
export URL="https://api.deepseek.com/v1"
export MODEL="deepseek-chat"

mkdir -p $RESULT_ROOT/$MODEL
export INCLUDE_MSG=no
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/nomsg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/nomsg.log
export INCLUDE_MSG=yes
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/msg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/msg.log

qwen-coder-plus
export KEY=""
export URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
export MODEL="qwen-coder-plus"

mkdir -p $RESULT_ROOT/$MODEL
export INCLUDE_MSG=no
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/nomsg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/nomsg.log
export INCLUDE_MSG=yes
python $EVAL_SCRIPT -d $DATASET_ROOT/test.jsonl -t online -m $MODEL -u $URL -k $KEY -o "$RESULT_ROOT/$MODEL/msg.jsonl" 2>&1 | tee $RESULT_ROOT/$MODEL/msg.log