(res)
| 183 | return total_results |
| 184 | |
| 185 | def cal_cmmlu(res): |
| 186 | print("\n\n\n") |
| 187 | results = {} |
| 188 | res = {k.split("-")[-1]: float(v) for k, v in res.items()} |
| 189 | for k, v in TASK_NAME_MAPPING.items(): |
| 190 | avg_acc = np.mean(list(map(lambda x: res[x], v))) |
| 191 | print("%s acc: %.4f " % (k, avg_acc)) |
| 192 | results[k] = "%.4f" % avg_acc |
| 193 | avg_all_acc = np.mean(list(res.values())) |
| 194 | print("AVERAGE acc:%.4f " % avg_all_acc) |
| 195 | results ['AVERAGE'] = "%.4f" % avg_all_acc |
| 196 | |
| 197 | return results |
| 198 | |
| 199 | def main(args): |
| 200 |