Question about text-code evaluation

[Zhao0278]

I trained the CodeBert model using CodeSearchNet on nl-code-search-webquery and fine-tuned it on Cosqa. But I have no access to the test_webquery.json file in the command from https://github.com/microsoft/CodeXGLUE/tree/main/Text-Code/NL-code-search-WebQuery#readme:

python code/run_classifier.py \
			--model_type roberta \
			--do_predict \
			--test_file test_webquery.json \
			--max_seq_length 200 \
			--per_gpu_eval_batch_size 2 \
			--data_dir ./data \
			--output_dir ./model_cosqa_continue_training/checkpoint-best-aver/ \
			--encoder_name_or_path microsoft/codebert-base \
			--pred_model_dir ./model_cosqa_continue_training/checkpoint-last/ \
			--prediction_file ./evaluator/webquery_predictions.txt 
			

I tried dev_webquery. json (from CodeSearchNet) and cosqa_dev. json. The validation results on cosqa_dev. json are similar to those in readme, but the accuracy of dev_webquery. json is very low. Additionally, the accuracy results in readme are inconsistent with those in the paper. Is the evaluation result in the paper obtained on CodeSearchNet or cosqa?