VideoChat2, 仅finetune stage3后，视频描述存在大量重复

[ruishuzhao]

作者好，
通过对videochat2中第三阶段的训练（训练集数据均来源于论文中提供的数据集），通过训练后模型对视频进行描述，会出现大量重复语句。
提问内容为："Describe the following video clip in detail."
回答如："The video clip shows a woman wearing a black shirt and black pants standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a"
其中反复出现“The woman is wearing a black shirt and black pants and is standing in a dark room. She is holding a white shirt and putting it into a washing machine. ”。

针对这种finetune后出现重复的问题，有比较好的解决方案吗？

辛苦帮忙解答一下。多谢~

[Andy1621]

这是一个不错的学术topic，因为在我们之前一些模型的推理上，也发现类似的问题，但具体是什么原因导致的不甚清楚。可以尝试修改不同的prompt，或者进行一些多轮对话提示，找一找原因或者解决方案

[ruishuzhao]

这是一个不错的学术topic，因为在我们之前一些模型的推理上，也发现类似的问题，但具体是什么原因导致的不甚清楚。可以尝试修改不同的prompt，或者进行一些多轮对话提示，找一找原因或者解决方案

再次感谢~

[yinanhe]

这种复读机问题在LLM中就存在，而且也有一些相应的介绍和探索
对于VideoChat中，您可以尝试调整answer函数中的num_beams, do_sample, top_p, repetition_penalty, temperature. 其中调整 repetition_penalty可能效果会大一点，但是也会影响生成文本的随机性和多样性。您可以尝试repetition_penalty的取值范围调整到1以上进行探索一下~

Ask-Anything/video_chat2/conversation.py

Lines 60 to 75 in 078540a

	def answer(self, conv, img_list, max_new_tokens=200, num_beams=1, min_length=1, top_p=0.9,
	repetition_penalty=1.0, length_penalty=1, temperature=1.0):
	conv.messages.append([conv.roles[1], None])
	embs = self.get_context_emb(conv, img_list)
	outputs = self.model.llama_model.generate(
	inputs_embeds=embs,
	max_new_tokens=max_new_tokens,
	stopping_criteria=self.stopping_criteria,
	num_beams=num_beams,
	do_sample=True,
	min_length=min_length,
	top_p=top_p,
	repetition_penalty=repetition_penalty,
	length_penalty=length_penalty,
	temperature=temperature,
	)