This application is used to track your day to day expenses using Natural Language. It uses 2 models for predicting expense details from natural language:
- Fine tuned Palm2 Bison (Hosted using Google API)
- Fine tuned Llama2-7B (Hosted locally).
The data for fine tuning was generated using GPT-3.5 with the following prompt:
Below is a input that describes an expense. Write a response in json format that appropriately completes the request.
Response is a json string with fields - account_type (CREDIT or DEBIT), category, sub_category, reason (Explain detailed reason if available), third_party - person who gave to got the money (Amount in Indian Rupees).
Generate appropriate response json string for the input expense. Response must be in only json string format strictly.
### Input:
I gave 5000 rupees to my friend for a personal loan repayment.
### Response:
With this technique, the 1000 data points was generated and both the models are fine tuned in following manner:
- Palm2 Bison was fine tuned on Google AI Studio Platform by importing the generated dataset with following configurations:
- Max Output Tokens: 256
- Temperature: 0.4
- Learning Rate: 0.02
- Batch Size: 16
- Epochs: 10
- Combined Loss: 0.01
- Llama2 was fine tuned using Ludwig AI and Transformers Framework on Tesla T4 Machine (Google Colab) with following configurations:
- Max Output Tokens: 256
- Temperature: 0.1
- Learning Rate: 0.0004
- Batch Size: 2
- Epochs: 10
- Combined Loss: 0.06