Finetune a model using peft adapters. You may use any python library like transformers and/or bits&bytes etc. PEFT offers parameter-efficient training methods. This method relies on a method called Low Rank Adaptoers (LoRA). Instead of fine-tuning the entire model you just have to fine-tune these adapters and load them properly inside the model.
Please use a small dataset like "Abirate/english_quotes" to finetune on a small model like "facebook/opt-125m".
- Fork or clone this repo
- Maintain the directory structure given for the project
- Use Python 3.7+
- If you need additional imports specify them in
requirements.txt
- Model should be trained on the given dataset.
- Create a model artifact and save it under
/models. - Report accuracy on
test.
- Serve the model as a REST API using FastAPI
- Be able to use CURL to send in text input and return the prediction.
The dataset you will be using contains ~2.51K rows of English quotes with author and tags. You need to finetune the base model to generate quotes.
You can split the Dataset into two splits train and test.
Timebox this challenge to 8-10 hours. After completing the assignment, please compress whole repo and send it.