nuggets icon indicating copy to clipboard operation
nuggets copied to clipboard

能否作为一个库发布

Open simplew2011 opened this issue 1 year ago • 2 comments

  • 能否作为一个pypi库发布
  • 【Predefined Task Set】,设置100,【待处理指令数据集】的每个样本都需要前向推理100次,速度较慢;有没有实验设置10个是不是能达到精度要求
  • 【Predefined Task Set】是否一定要从【待处理指令数据集】里选取
  • 示例参数里--prompt_path "datasets/alpaca_gpt4/alpaca_gpt4_data.json" --test_path "datasets/alpaca_gpt4/alpaca_gpt4_kmeans_100.json",【待处理指令数据集】被定义为prompt_path,【Predefined Task Set】被定义为test_path,交换一下参数定义是不是更好理解
  • tasks/alpaca.py#L81,这里prefix为什么为空呢
  • 默认参数跑,为什么win_rate和norm_win_rate的值是一样的

simplew2011 avatar Aug 05 '24 08:08 simplew2011

--model_type local --model_path /data1/xhj/llm/Llama-2-7b-hf

ValueError: Trying to set a tensor of shape torch.Size([32000, 4096]) in "weight" (which has shape torch.Size([32001, 4096])), this looks incorrect.

simplew2011 avatar Aug 06 '24 10:08 simplew2011

  • 切换transformers==4.28.1,使用daryl149/llama-2-7b-chat-hf,预测发现结果和alpaca_gpt4_sorted_score_sample_100_kmeans.json里的socre不对应

  • --prompt_path alpaca_gpt4_sorted_score_sample_100_kmeans.json --test_path "datasets/alpaca_gpt4/alpaca_gpt4_kmeans_100.json"

image

simplew2011 avatar Aug 06 '24 13:08 simplew2011