DeepSeek-Math/evaluation/configs/zero_shot_test_configs.json

34 lines
1.1 KiB
JSON
Raw Normal View History

2024-02-06 02:27:40 +00:00
{
"gsm8k-test": {
"test_path": "datasets/gsm8k/test.jsonl",
"language": "en",
"tasks": ["tool", "cot"],
"process_fn": "process_gsm8k_test",
"answer_extraction_fn": "extract_last_single_answer",
"eval_fn": "eval_last_single_answer"
},
"math-test": {
"test_path": "datasets/math/test.jsonl",
"language": "en",
"tasks": ["tool", "cot"],
"process_fn": "process_math_test",
"answer_extraction_fn": "extract_math_answer",
"eval_fn": "eval_math"
},
"mgsm-zh": {
"test_path": "datasets/mgsm_zh/mgsm_zh.jsonl",
"language": "zh",
"tasks": ["tool", "cot"],
"process_fn": "process_mgsm_zh",
"answer_extraction_fn": "extract_last_single_answer",
"eval_fn": "eval_last_single_answer"
},
"cmath": {
"test_path": "datasets/cmath/test.jsonl",
"language": "zh",
"tasks": ["tool", "cot"],
"process_fn": "process_cmath",
"answer_extraction_fn": "extract_last_single_answer",
"eval_fn": "eval_last_single_answer"
}
}