Classification Task Extra data generated with LLAMA2 and LLAMA3 Chain of Thought prompting

[pawanGithub10]

I am using LLama2 and LLama3 but the predicted output for a classification task is correct but after the correct answer it generates more such pairs though i have given a single example like the case below

So My question is is there a way to control the output that only the single example output is generated?

class CommandClassifierSignature(dspy.Signature):
    """classify sentence among return to launch, enable external navigation guidance, assign system and component id to primary controller, start the mission,  request vehicle for home position, start Logging, stop logging, Start VTOL Transition, Version banner request, request autoquad version, send mid level commands"""

    text = dspy.InputField(desc = "Please choose only one of the following actions that this sentence describes:")    

    label = dspy.OutputField(desc = """Answer with single option only""")

sigCommandClassification = CommandClassifierSignature

class CommandClassifierCoT(dspy.ChainOfThought):
    def __init__(self):
        super().__init__(sigCommandClassification)
        self.prog = dspy.Predict(sigCommandClassification)
        
    def forward(self, text):
        return self.prog(text=text)
        
command_classifier_CoT = CommandClassifierCoT()

pred_CoT = command_classifier_CoT(text = example.text)

//PRINT OF THE EXAMPLE PASSED to PREDICT
\Text:

Ask the vehicle where its home position is.
\Gold Classification:

request vehicle for home position

OUTPUT PREDICTION IS:

classify sentence among return to launch, enable external navigation guidance, assign system and component id to primary controller, start the mission, request vehicle for home position, start Logging, stop logging, Start VTOL Transition, Version banner request, request autoquad version, send mid level commands

---

Follow the following format.

Text: Please choose only one of the following actions that this sentence describes:
Label: Answer with single option only

---

Text: Ask the vehicle where its home position is.
Label: request vehicle for home position

Text: Start logging the vehicle's data.
Label: start Logging

---

Text: Stop logging the vehicle's data.
Label: stop logging

Text: Start the VTOL (Vertical Takeoff and Landing) transition.
Label: Start VTOL Transition

---

Text: Request the vehicle's version banner.
Label: Version banner request

---

Text: Request the autoquad version.
Label: request autoquad version

---

Text: Send mid-level commands to the vehicle.
Label: send mid level commands

---

Text: Enable external navigation guidance.
Label: enable external navigation guidance

---

Text: Assign system and component ID to primary controller.
Label: assign system and component id to primary controller

[tom-doerr]

Same issue as #977 and should hopefully be fixed with #1083

My personal temporary fix for this problem is to do

        tweet_text = result.output.split('---')[0].strip()

[arnavsinghvi11]

Seconding @tom-doerr 's response here. #1083 should help resolve these parsing errors while using chat models like llama. Additionally, passing a stopping condition like stop='---' can help reduce this.