Zhe Chen
Zhe Chen
感觉想要finetune coco caption涨点太难了,咋tune都掉点...
现在这两个问题还有吗
> 请问下正负样本的构造必须要加`\n` 前缀吗? 关于这块儿有没有其他说明呢 @czczup 所有有图像的数据都需要在第一轮对话的question里加上`\n`,如果是纯文本数据不要加这个前缀。
That's great, thank you very much!
需要使用OpenAI的key,用GPT4辅助抽取答案,37%的点数是因为没有使用GPT4
参考这部分的文档:https://internvl.readthedocs.io/en/latest/internvl1.2/evaluation.html#mathvista
Hi, we may release the annotation files in the JSONL format that we use. However, to make them usable for everyone, we will need to create a document detailing the...
Hi, thanks for your attention. Finetuning a 34B LLM with 32GB GPUs might be somewhat challenging. You could consider trying to only fine-tune the MLP connector or Lora in the...
> 我也遇到了同样的问题,我是有8张v100,PER_DEVICE_BATCH_SIZE设为1,BATCH_SIZE设为8 您好,您可以先试试只微调MLP层,看看会不会OOM ```shell --freeze_llm True \ --freeze_mlp False \ --freeze_backbone True \ ```
噢噢,可能需要在shell上面加一行: export LAUNCHER=pytorch 麻烦您试试看