Sentences

Utilities for sentence-level text segmentation and tokenization tailored to LLM tokenizers.

This package is designed to support sentence-level (“Thought Anchor”) analyses like those in:

Bogdan, P. C.*, Macar, U.*, Nanda, N.°, & Conmy, A.° Thought anchors: Which LLM reasoning steps matter? 2025. https://arxiv.org/abs/2506.19143.

Features

• Splits a given text into sentences
• Avoids common issues (e.g., "Dr. Fu" shouldn't be split into two sentences)
• Respects standard LLM tokenization patterns (e.g., leading-space tokens)
• Given a tokenizer, returns token ranges for each sentence in the tokenized input text

Installation

pip install sentences

Sentence Splitting

Sentences split here adhere to typical LLM tokenization strategies. For example, this sentence "I love my cat. It is big." should be split with a leading space rather than a trailing one, ["I love my cat.", " It is big."]

from sentences import split_text_to_sentences

text = "Dr. Smith went to the store. They bought some milk. It cost $3.50."
sentences, positions = split_text_to_sentences(text)

for i, (sent, pos) in enumerate(zip(sentences, positions)):
    print(f"{i}: {repr(sent)}")
    # 0: 'Dr. Smith went to the store.'
    # 1: ' They bought some milk.'
    # 2: ' It cost $3.50.'

Token Range Extraction

You can use this package to get the exact token ranges for sentences. You can use this to split up a model's chain-of-thought into sentences. You can include pre_string, where you provide a string that will appear before your sentences (e.g., a chat template), and the token ranges will respect that.

Token ranges are calculated by repeatedly appending a new sentence to the pre_string, tokenizing the new string, and counting the number of tokens. This helps avoid tokenization oddities. Simply tokenizing each sentence independently can cause problems.

from sentences import get_token_ranges
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-14B-Instruct")

# Example with Qwen-2.5 chat template
pre_string = """<|im_start|>system
This system message is just for demonstration purposes.<|im_end|>
<|im_start|>user
Solve this math problem step by step.<|im_end|>
<|im_start|>assistant
<think>

"""

sentences = ["Let me think about this problem.", " First, I'll break it down."]
ranges = get_token_ranges(sentences, tokenizer, pre_string)
tokens_all = tokenizer.batch_decode(tokenizer.encode(pre_string + ''.join(sentences)))

for sent, (start, end) in zip(sentences, ranges):
    print(f"Tokens [{start}:{end}] = '{sent}'\n\t{tokens_all[start:end]}")
    # Tokens [39:46] = 'Let me think about this problem.'
    #   [' Let', ' me', ' think', ' about', ' this', ' problem', '.']
    # Tokens [46:54] = ' First, I'll break it down.'
    #   [' First', ',', ' I', "'ll", ' break', ' it', ' down', '.']

Note on CoT pre-filling

gpt-oss models don't use <think> tags but instead employ a special format:

pre_string = """<|start|>system<|message|>You are ChatGPT, a large language model trained by OpenAI.
Knowledge cutoff: 2024-06
Current date: 2025-11-22<|end|><|start|>user<|message|>Solve this math problem step by step.<|end|><|start|>assistant<|channel|>analysis<|message|>"""