-
Notifications
You must be signed in to change notification settings - Fork 1.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
adding peft lora example notebook for ner #2126
Conversation
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Context: #2119
Thanks @JINO-ROHIT for taking this task. Good find of a base notebook to adapt. @HarikrishnanK9 is this what you were looking for?
I haven't done a fine-grained analysis yet, but it looks like the compute_metrics
are note being used. Therefore, we don't get the interesting metrics defined in there. Could you please check that?
will do |
i added the eval metric part, the recall and precision seem quite low, seems to be the classes are imbalanced, would you like me to handle that? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks a lot for the updates. I checked the notebook and found the source for the low performance, namely the learning rate. Check my comment.
Also, when I first ran this notebook, I got an error:
ImportError: To be able to use evaluate-metric/seqeval, you need to install the following dependencies['seqeval'] using 'pip install seqeval' for instance'
So let's mention at the beginning which packages the user should install. We can assume that torch and PEFT (and all their dependencies) are installed, but besides seqeval
, let's also mention datasets
and evaluate
.
Moreover, it would be nice to set the batch size to a higher value. The base model is quite small, so even with batch sizes of 32 or more, most users should be able to run the notebook. This will speed up training considerably. You can add a comment that users have to reduce the value if they run OOM.
], | ||
"source": [ | ||
"# Configure LoRA (Low-Rank Adaptation) for fine-tuning\n", | ||
"peft_config = LoraConfig(target_modules = [\"query\", \"key\"])\n", |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
"peft_config = LoraConfig(target_modules = [\"query\", \"key\"])\n", | |
"peft_config = LoraConfig(target_modules=[\"query\", \"key\"], task_type=TaskType.TOKEN_CLS))\n", |
TaskType
needs to be imported from PEFT.
" \"bert-finetuned-ner-lora\",\n", | ||
" evaluation_strategy=\"epoch\",\n", | ||
" save_strategy=\"epoch\",\n", | ||
" learning_rate=2e-5,\n", |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The learning rate should be higher for LoRA. This is probably why you find worse results than expected. I checked 2e-4, 2e-3, and 2e-2, and got best results for 2e-3. However, they were still slightly worse than full fine-tuning and I only ran for 3 epochs, so some further experiments could yield better results still.
"metadata": {}, | ||
"outputs": [], | ||
"source": [ | ||
"# Import required libraries\n", |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Let's add a few sentences to the beginning before starting with the code. Something along the lines of: "In this notebook, we learn how to ...". You can also add the link to the course here.
oh yeah, i totally missed the lr part. The metrics look waaay better now, thanks! |
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for adding this notebook to show NER with PEFT, LGTM.
From my point of view, it's ready to merge. I'll give @HarikrishnanK9 some time to take a look to possibly give more feedback. If we don't hear back by Friday, I'll merge.
gotcha |
@JINO-ROHIT Thank You @BenjaminBossan Just go ahead with the merge. |
Hi @BenjaminBossan ,
Ive added a lora example for NER, which is mainly a clone of the huggingface NER notebook by Sylvain here https://huggingface.co/learn/nlp-course/chapter7/2#data-collation with couple changes here and there.
lmk what you think