language:

• en
  task_categories:
• text-generation
  tags:
• synthetic
• conversational
• reasoning
• chain-of-thought
  size_categories:
• 1K<n<10K
  configs:
• config_name: default
  data_files:
  • split: train
    path: train.parquet
    dataset_info:
    features:
  • name: source
    dtype: string
  • name: category
    dtype: string
  • name: messages
    list:
    • name: role
      dtype: string
    • name: content
      dtype: string
    • name: reasoning
      dtype: string
      splits:
  • name: train
    num_examples: 1000

reasonable-chats

Synthetic multi-turn conversations between a user and a helpful assistant, where each
assistant turn carries an explicit reasoning trace whose length scales with the
difficulty of the question. Short questions get brief traces; complex ones get
step-by-step deliberation.

The conversations are generated by google/gemma-4-31b-it via structured output, seeded
by prompts from Dahoas/synthetic-hh-rlhf-prompts.

Fields

• source — SHA-256 hash of the seed prompt the conversation was generated from.
• category — the topic type: one of factual, explanation, how_to, math, coding, writing, reasoning.
• messages — the conversation as an alternating list of turns, starting with user and ending with assistant:
  • role — user or assistant.
  • content — the message text.
  • reasoning — the assistant's private thinking trace for that turn; null for user messages.

Generation and filtering

Each seed prompt is first passed through a gate. Topics are rejected when a good answer
would require time-sensitive specifics (recent events, or any calendar year from 2000
onward), would make the assistant refer to itself or claim to be an AI, or would call for
a refusal or a "consult a professional" caveat. Accepted topics produce one to three
conversations of at least two turns each.

Generated conversations are then filtered. A conversation is dropped entirely if it leaks
assistant-identity or refusal language, defers the user to an expert, mentions a year
between 2000 and 2040, or opens its reasoning by restating the request. Markdown emphasis
and em/en-dashes are stripped from the text.

Loading

from datasets import load_dataset

dataset = load_dataset("breitburg/reasonable-chats", split="train")