| """ Extract the prediction samples from the raw prediction file. | |
| i.e., ${data}/test_raw.jsonl to ${data}/test.jsonl | |
| """ | |
| import json | |
| from jsonargparse import CLI | |
| def context2input(l, start=0): | |
| """ Given a list of context, just following UniCRS to format the context. | |
| """ | |
| context = '' | |
| for i, utt in enumerate(l): | |
| if utt == '' or utt == ' ': | |
| continue | |
| if i % 2 == start: | |
| context += '\nUser: ' | |
| else: | |
| context += '\nSystem: ' | |
| context += utt | |
| return context + '\nSystem: ' | |
| def main(from_json: str = None, to_json: str = None, start: int = 0): | |
| with open(from_json) as fr, open(to_json, 'w') as fw: | |
| for l in fr: | |
| l = json.loads(l) | |
| if l['is_user'] == 0 and len(l['rec']) > 0: | |
| l['input'] = context2input(l['context'], start=start) | |
| fw.write(json.dumps(l) + '\n') | |
| if __name__ == '__main__': | |
| # python extract_prediction.py --from_json=inspired/test_raw.jsonl --to_json=inspired/test.jsonl --start 0 | |
| # python extract_prediction.py --from_json=redial/test_raw.jsonl --to_json=redial/test.jsonl --start 1 | |
| # python extract_prediction.py --from_json=reddit/test_raw.jsonl --to_json=reddit/test.jsonl --start 1 | |
| CLI(main) | |