How to reduce repetition with frequency_penalty

The frequency_penalty parameter helps you control how often the model repeats words or phrases.

Positive values → penalize tokens that have already appeared → model is less likely to repeat itself.
Negative values → encourage more repetition (rarely used).

How it works:

After each token is generated, the model adjusts the probability of generating the same tokens again, based on how often they have already been used.
This helps you create more varied and natural-sounding text.

Parameter details:

Parameter	Type	Range	Default
`frequency_penalty`	Double	-2.0 to 2.0	0.0

When to use `frequency_penalty`:

Scenario	Suggested Value
Model is repeating itself too much	0.5 to 1.0
Long-form content, essays, articles	0.5 to 1.0
Dialogue-heavy apps (avoid echoing user)	0.5 to 1.0
Want creative, varied phrasing	0.5 to 1.5

1 # Install SarvamAI
2 !pip install -Uqq sarvamai
3 from sarvamai import SarvamAI

1 # Initialize the SarvamAI client with your API key
2 client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")

1 # Example: Apply frequency_penalty to avoid repetition
2 response = client.chat.completions(
3     messages=[
4         {"role": "system", "content": "You are a travel blogger assistant."},
5         {"role": "user", "content": "Write a paragraph about the beaches of Goa."}
6     ],
7     frequency_penalty=1.0  # Reduce repetitive phrases
8 )

1 # Receive assistant's reply as output.
2 print(response.choices[0].message.content)

How it works:

Parameter details:

When to use frequency_penalty:

When to use `frequency_penalty`: