perplexity 微调之后变大了？需要继续微调？

shibing624 / MedicalGPT

MedicalGPT: Training Your Own Medical GPT Model with ChatGPT Training Pipeline. 训练医疗大模型，实现了包括增量预训练(PT)、有监督微调(SFT)、RLHF、DPO、ORPO。

Apache License 2.0

3.38k stars 508 forks source link

perplexity 微调之后变大了？需要继续微调？ #416

Open cqray1990 opened 3 months ago

cqray1990 commented 3 months ago

微调之前估计效果： eval metrics eval_loss = 2.5866 eval_runtime = 0:02:10.44 eval_samples = 1000 eval_samples_per_second = 7.666 eval_steps_per_second = 0.958 perplexity = 13.2845

sft 微调之后的估计效果，perplexity 怎么还变大了

eval metrics epoch = 2.0 eval_loss = 2.8984 eval_runtime = 0:00:00.33 eval_samples = 10 eval_samples_per_second = 29.506 eval_steps_per_second = 2.951 perplexity = 18.1445

shibing624 commented 2 months ago

建议： 1.训练集质量不高，提升数据质量； 2.lr 调低 3.batch size调大