Update README.md

README.md

@@ -35,8 +35,7 @@ tokenizer = AutoTokenizer.from_pretrained("paytm/text2graph1.5b-original")
 model = model.to(device)
 
 # Prepare your input text
-text = """
-''
+text = '''
 {
   "task": "Extract structured information from SMS messages to identify financial details, message purpose, and related information",
   "instructions": [
@@ -71,16 +70,17 @@ text = """
     "key_themes: Extract only 3-5 main themes actually discussed in the SMS as an array of strings - avoid redundant or fabricated themes",
     "message_type: Categorize as exactly one of: alert message, delayed payment, repayment reminder, auto-pay reminder, refund information, payment failed information, settlement notification by payment aggregator, loan_emi_reminder, emi_confirmation, loan_offer, new_merchant_acquisition, product_offer",
     "product_type: Identify as one of: credit card, debit card, home loan, vehicle loan, merchant_qr_kit, soundbox (or null if not mentioned)",
-    "aggregator_name: Extract payment aggregator name like paytm, phonepe, gpay , bharatpe, one 97 communications (or null if not mentioned). The namees can be written with differen caps lock , inside urls etc. Pay attention and extract it correctly.",
+    "aggregator_name: Extract payment aggregator name like paytm, phonepe, gpay , bharatpe, one 97 communications (or null if not mentioned). The namees can be written with differen caps lock , inside urls etc. Pay attention and extract it correctly.Do not return bank names in aggregators",
     "bank_nbfc_fintech_name: Extract the name of the financial institution the message is from (or null if not present). A bank might have aggregator service but dont confuse the bank with aggregator",
     "emi_amount: Extract only a clearly identified EMI amount as a float number (or null if not mentioned)",
     "Total_overdue_amount: Extract only a clearly identified total overdue or bill amount as a float (or null if not mentioned)",
-    "available_bank_balance: Extract only a clearly identified available bank balance as a float (or null if not mentioned)",
+    "available_bank_balance: Extract only a clearly identified available bank balance as a float (or null if not mentioned). Do not mention available balance when not given",
     "min_overdue_amount: Extract only a clearly identified minimum amount due as a float (or null if not mentioned)",
     "overdue_date: Extract only a clearly identified payment due date as string (or null if not mentioned)",
     "collection_team_lead: Extract only explicit mentions of collection team involvement (or null if not present)",
     "legal_notice: Extract only explicit mentions of legal notices (or null if not mentioned)",
-    "customer_name: Extract only an explicitly identified customer name (or null if not mentioned)"
+    "customer_name: Extract only an explicitly identified customer name (or null if not mentioned)",
+    "settlement_amount": Amount credited by the aggregator to the bank account of the merchant/customer,
   ],
 **example**
 <input>"Dear Qagent, your XYZ credit card bill of Rs. 24,500 is due on 15-Jul-2023. Min amount due: Rs. 2,450. Account balance on 3-Jul-2023 is Rs 5,000.Pay now to avoid late fees. - XYZ Bank"</input>
@@ -97,19 +97,19 @@ text = """
   "overdue_date": "2023-07-15",
   "collection_team_lead": null,
   "legal_notice": null,
-  "customer_name": "Qagent"
+  "customer_name": "Qagent",
+  "settlement_amount":null
 }</output>
-
-<input>"Dear SBI Cardholder, outstanding of Rs. 9491.00, on your credit card ending 2852 is due on 31-JAN-25. Min. Amount Due: Rs. 475.00. Please ignore if already paid. Click here: https://sbicard.com/quickpaynet to make online payment."</input>
+<input>"Dear Customer, INR 9,307.00 credited to your A/c No XX1644 on 12/03/2025 through NEFT with UTR AXNPN07162650312 by PHONEPE PRIVATE LIMITED-PAYMENT AGGREGATOR ESCROW ACCOUNT, INFO: BATCHID:0010 F09 M22YWRZNKJE4SREDEMPTION-SBI"</input>
 <output>
-"""
+'''
 
 # Tokenize input and generate output
 inputs = tokenizer(text, return_tensors="pt")
 outputs = model.generate(**inputs.to("cuda"), max_new_tokens=185,do_sample=False,top_k=None,temperature=1.0,top_p=None)
 
 # Decode the output
-result = tokenizer.decode(outputs[0], skip_special_tokens=True)
+result = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
 print(result)
 ```