This is an experimental model fine-tuned on various balance sheets collected from financial services. The fine-tuning process was designed to adapt the TAPAS model to handle large numeric values and complex financial data structures commonly found in balance sheets.
How to Get Started with the Model
Use the code below to get started with the model.
# !pip install sugardata
from transformers import TapasTokenizer, TapasForQuestionAnswering
from sugardata.utility.tapas import generate_financial_balance_sheet, get_real_tapas_answer
# generate a financial balance sheet and ask a question
table = generate_financial_balance_sheet()
question = "What was the reported value of Total Debt in 2021?"
# load the model and tokenizer
model_name = "yeniguno/tapas-base-wtq-balance-sheet-tuned"
model = TapasForQuestionAnswering.from_pretrained(model_name)
tokenizer = TapasTokenizer.from_pretrained(model_name)
inputs = tokenizer(table=table, queries=[question], padding="max_length", return_tensors="pt")
# get the answer
answer = get_real_tapas_answer(table, model, tokenizer, inputs)
# 8873000.0
Training Details
- Epoch [1/5] Train Loss: 0.1514 Val Loss: 0.0107
- Epoch [2/5] Train Loss: 0.0135 Val Loss: 0.0098
- Epoch [3/5] Train Loss: 0.0116 Val Loss: 0.0081
- Epoch [4/5] Train Loss: 0.0081 Val Loss: 0.0071
- Epoch [5/5] Train Loss: 0.0049 Val Loss: 0.0043