Automotive_NER

This is my code for the Predii Automotive NER assignment.

Installing the required libraries

pip install -r requirements.txt

If you face any issues installing the libraries with requirements.txt, use the following commands -

pip install einops datasets bitsandbytes accelerate peft flash_attn
pip uninstall -y transformers
pip install git+https://github.com/huggingface/transformers
pip install --upgrade torch
pip install replicate
pip install openai
pip install jsonschema

Generating Artificial Data

Generate labeled data artificially using OpenAI gpt-3.5-turbo LLM. Before generating the data, set the OpenAI_API_Key as an environment variable. You can generate the OpenAI_API_Key from https://platform.openai.com/api-keys.

os.environ["OPENAI_API_KEY"] = <YOUR_OPENAI_API_KEY>

To generate the data, run datag_gen.py -

python data_gen.py

This will save the data in data/gpt_gen_data.csv. You can change the name of the csv file in data_gen.py.

Fine-tuning

We fine-tune using the Low-Rank Adaptation (LoRA) technique to get better performance with optimized memory utilization on A100 Nvidia GPU. We use flash attention which requires GPUs that are based on the Ampere architecture.

Fine-tune the Microsoft Phi-2 model by running train.py. Before running this train.py, set the training configs in configs/config.json.

python train.py

Generate a HuggingFace access token and set it in configs/config.json -

"hf_access_token": <YOUR_HF_ACCESS_TOKEN>

Evaluation

We evaluate the LlaMa 2 7b, Phi-2 and the fine-tuned Phi-2 models on data/test.csv. For evaluation, run evaluate.py after setting the Replicate API token as an environment variable. You can generate the Replicate_API_Token from https://replicate.com/account/api-tokens

os.environ["REPLICATE_API_TOKEN"] = <YOUR_REPLICATE_API_TOKEN>

Testing The Model On Your Text

Test the model on your text by running the test.py file -

python test.py --text <YOUR_TEXT>