2B model outputting math problem solving CoT instead of entities
#2
by xbill - opened
Hello, I am trying out the 2B model on Mac with the following prompt (generated by the tokenizer)
<|im_start|>system
You are NuExtract, an information extraction tool created by NuMind.<|im_end|>
<|im_start|>user
# Template:
{"name": "[string]", "company": "[string]", "amount": "[string]", "date": "[string]"}
# Context:
Hello John, the invoice for $500 from ABC Company is due on Friday.<|im_end|>
<|im_start|>assistant
I am expecting the list of entities as output, but here's what the model generates
[
" John has already paid $200. What is the balance due on the invoice?\nAnswer Choices: (A) 300 (B) 200 (C) 100 (D) 500 (E) 400\n\nTo find the balance due on the invoice, we subtract the amount John has already paid from the total invoice amount. Given that the invoice is for $500 and John has already paid $200, the balance due is:\n\n\\[ \\text{Balance due} = \\text{Invoice amount} - \\text{Payment amount} \\]\n\\[ \\text{Balance due} = 500 - 200 \\]\n\\[ \\text{Balance due} = 300 \\]\n\nTherefore, the correct answer is (A) 300. John is still due $300 on the invoice. However, the question seems to be a mix of a mathematical problem and a word problem, which might not be directly related. If the context is meant to be a mathematical problem, the correct answer is 300. If it's a word problem, the correct answer is not provided in the options. The question might need clarification to match the context accurately."
]
Am I doing something wrong? I have used the code snippets provided in the README. BTW, the AutoModelForVision2Seq class seems to be deprecated in favor of AutoModelForImageTextToText, but I preferred to stick with the old one.
Thanks in advance for your help