Bad performance for three generation modes.

○ kb：{'rouge1': 0.10795632401934566, 'rouge2': 0.005502676702905536, 'rougeL': 0.10485264583478511, 'rougeLsum': 0.10448227546441474, 'bert_score_precision': 0.5239911143481731, 'bert_score_recall': 0.5586187492311001, 'bert_score_f1': 0.537801007553935, 'mem_cost': 22861053952
○ ICL-50个：{'rouge1': 0.10135071880712321, 'rouge2': 0.06694877154125624, 'rougeL': 0.09862427623499491, 'rougeLsum': 0.09862427623499491, 'bert_score_precision': 0.32884539023041726, 'bert_score_recall': 0.4558024096488953, 'bert_score_f1': 0.3758223417401314, 'mem_cost': 25.775390625}
○ zero-shot：{'rouge1': 0.030773897832609545, 'rouge2': 0.0015176552801414067, 'rougeL': 0.028298936576585366, 'rougeLsum': 0.027872349274998064, 'bert_score_precision': 0.3853674686700106, 'bert_score_recall': 0.4561145968735218, 'bert_score_f1': 0.4070150001347065, 'mem_cost': 30863785984}

use enron as train set and synthetic as eval set

args.command is "generation".

niter=900, loss=1.565

base mode is Meta-Llama-3-8B

the performance of ICL is also very bad ??? kb_size is 50. OOM when kb_size=80.

Start at the generation entry point selected by args.command="generation" and reproduce the three modes with Enron as training data and synthetic as evaluation data. Compare the reported ROUGE, BERTScore, and memory values for niter=900, including the kb_size=50 result and kb_size=80 OOM. Done requires identifying the cause of the poor performance or memory failure and documenting a verified correction.