Limit the number of characters DeepSeek R1 can use for thinking.

thinking-cap.py

import outlines
from transformers import AutoTokenizer

model_string = 'deepseek-ai/DeepSeek-R1-Distill-Qwen-7B'
# model_string = 'deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B' # For small machines

model = outlines.models.transformers(
    model_string,
    device='cuda', # also 'cpu', 'mps','auto'
)
tokenizer = AutoTokenizer.from_pretrained(model_string)

thinking_regex = r'<think>(.|\n){500}\n\[THINKING_TRUNCATED\]\n</think>(yes|no)'

prompt = tokenizer.apply_chat_template(
    [
        {'role': 'system', 'content': 'You are a helpful assistant.'},
        {'role': 'user', 'content': 'Roses are red. Violets are blue. Are roses and violets the same color? Yes or no.'},
    ],
    tokenize=False,
    add_generation_prompt=True,
)

# Generator 
generator = outlines.generate.regex(model, thinking_regex)
print("Generator created")

# Result 
result = generator(prompt)
print(result)

Example response:

<think>
Okay, so I came across this question: "Roses are red. Violets are blue. Are roses and violets the same color?" At first glance, it seems straightforward, but I want to make sure I understand it properly. Let me break it down.

First, the statements "Roses are red" and "Violets are blue" are pretty common. I know that roses do come in various colors, but the most common ones, like the ones you get in bouquets, are red. Similarly, violet flowers, like the실, are often depicted as blue in color, at
[THINKING_TRUNCATED]
</think>yes