search

TUDB-Labs / MixLoRA

State-of-the-art Parameter-Efficient MoE Fine-tuning Method
Apache License 2.0
51 stars 7 forks source link

Missing key in mlora.json when training and evaluating multi-task #3

Closed Edmund-Lai closed 1 month ago

Edmund-Lai commented 1 month ago

After executing python ./launch.py gen --template mixlora --tasks "arc-c;arc-e;boolq;obqa;piqa" --multi_task True --adapter_name mixlora to generate mlora.json, I tried python ./launch.py run --base_model <Path to Your Base Model> for training and evaluating on multi-task setting.

Then I got the following error message:

Traceback (most recent call last):
  File ".../mlora/mlora.py", line 268, in <module>
    mlora.train(mlora.Dispatcher(config, tokenizer), model,
  File ".../mlora/mlora/dispatcher.py", line 201, in __init__
    dataload_function=lora["dataloader"],
KeyError: 'dataloader'

I checked the code in dispatcher.py, the dataloader should be a Callable object. How can I fix this error?

Thanks.

mikecovlee commented 1 month ago

Which version of m-LoRA you are using with?

Edmund-Lai commented 1 month ago

Which version of m-LoRA you are using with?

0.3.2, same as the README mentioned.

mikecovlee commented 1 month ago

All our experiments were conducted under version 0.3.2. Based on your description, I cannot reproduce this problem. There may be some problems with your Git repository. You can try using the latest version. If the problem persists, you can add me on WeChat (ID: mikecovlee) to discuss in detail.

mikecovlee commented 1 month ago

Closing issue because no further response.