Datasets:

jackwarner
/

multi-turn-conversations-v2

license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - code
  - math
  - music
pretty_name: Multi-Turn Conversations and Prompting for Different Answers (v2)
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: ID
      dtype: int64
    - name: P1
      dtype: large_string
    - name: R1
      dtype: large_string
    - name: P2
      dtype: large_string
    - name: R2
      dtype: large_string
    - name: P3
      dtype: large_string
    - name: R3
      dtype: large_string
    - name: P4
      dtype: large_string
    - name: R4
      dtype: large_string
  splits:
    - name: train
      num_bytes: 369478
      num_examples: 110
  download_size: 202526
  dataset_size: 369478

Multi-turn Prompts Dataset (v2)

Description

This dataset consists of multi-turn, text-only conversations in English designed to analyze creativity and persistence in large language models across repeated follow-up prompts. The conversations were generated using gpt-5-nano (served through OpenRouter as openai/gpt-5-nano) and structured to systematically probe how an LLM reformulates, elaborates on, or diverges from prior responses over multiple turns.

The dataset focuses on controlled conversational dynamics rather than task completion accuracy. Each conversation begins with a human-authored initial prompt and is followed by a sequence of archetype-driven follow-up prompts intended to elicit varying degrees of creativity, reformulation, clarification, and persistence in the model’s responses.

The data was generated in October 2026 and is intended for research use in analyzing multi-turn behavior, semantic novelty, and response consistency in large language models.

Definition of Creativity

In this dataset, creativity is assessed turn by turn: each response is compared against what was already said earlier in the same conversation. A response expresses creativity through two outlets:

Novelty: the introduction of new content, meaning ideas, examples, framings, or information that did not appear in any earlier turn of the conversation.
Recombination: the unique recombination of ideas that were already mentioned, for example connecting earlier points in a new way, reframing them through analogy or a different perspective, or synthesizing them into a new structure.

A response's creativity is therefore based on how much novel content it introduces as well as how uniquely it recombines what was previously mentioned. A response that simply restates prior content, adding little that is new and reorganizing little of what exists, is considered persistent rather than creative.

Key Features

Multi-turn conversations with up to four prompt–response pairs (P1–R1 through P4–R4).
Follow-up prompts are drawn from 10 distinct archetypes, each designed to stress different conversational behaviors.
Archetypes are applied consistently across conversations to enable controlled comparison of response changes across turns.
Suitable for sentence-level or turn-level analysis of creativity (novelty and recombination), semantic similarity, and persistence across dialogue turns.

Changes from v1

This version replaces 9 of the 11 initial (P1) prompts of jackwarner/multi-turn-conversations with new starter prompts. Two prompts are carried over unchanged: "How did each orchestra instrument develop?" and "What is the history of the Roman Empire?". The 10 follow-up archetypes are unchanged. Responses were regenerated through OpenRouter's Chat Completions API rather than OpenAI's Responses API, so default generation behavior may differ slightly from v1.

Responses in this version are also length-limited: every conversation begins with the system prompt "Respond in at most 10 sentences. If you use a list, count each item as a sentence.", which stays in the conversation history for all four turns. Responses are therefore shorter than v1's, which were generated without a length limit.

Dataset Source

The initial prompts were authored by the dataset creator in collaboration with members of the DRAGN Lab at Brigham Young University (BYU).
Follow-up prompts were programmatically selected from predefined archetypes, and responses were generated using gpt-5-nano via the OpenRouter Chat Completions API, with a system prompt limiting each response to at most 10 sentences.

This dataset was created solely for research and analysis purposes and does not contain human conversational transcripts.

Structure & Fields

The dataset is organized in a wide, single-row-per-conversation format with the following fields:

P1, R1: The initial user prompt and the model’s first response.
P2, R2: The first archetype-driven follow-up prompt and corresponding response.
P3, R3: The second archetype-driven follow-up prompt and corresponding response.
P4, R4: The third archetype-driven follow-up prompt and corresponding response.

Each row represents a complete multi-turn interaction, allowing direct comparison of how the model’s responses evolve across successive prompts.

Follow-up Prompt Archetypes

Each conversation applies one of the following 10 archetypes, with three follow-up prompts used for turns P2, P3, and P4:

Reformulation / Perspective / Constraint
Encourages rewording, metaphorical explanation, and constrained restatement.
Elaboration / Example / Verification
Requests expanded explanations, concrete examples, and verification of semantic consistency.
Clarification / Rephrasing / Clarification
Simulates user confusion and repeated clarification requests.
Example / Variation / Constraint
Pushes the model to generate diverse examples and identify shared structure.
Perspective / Elaboration / Verification
Shifts explanatory perspective and reconnects to the original idea.
Constraint / Example / Reformulation
Combines concise explanation, exemplification, and expanded restatement.
Challenge / Refine / Summarize
Introduces counterarguments and refinement pressure.
User Confusion
Mimics repeated expressions of misunderstanding.
User Repetition
Models dissatisfaction and repeated requests for alternative answers.
User Correction / Confusion
Challenges correctness and coherence of prior responses.

These archetypes are designed to systematically probe how models balance novelty and persistence across turns.

Intended Use Cases

This dataset is intended for:

Studying creativity, as both novelty (new content) and recombination (unique reuse of prior content), in multi-turn LLM responses.
Analyzing persistence vs. divergence in conversational agents.
Evaluating response stability under repeated reformulation or user dissatisfaction.
Research in dialogue modeling, conversational consistency, and adaptive generation strategies.

It is particularly suitable for embedding-based similarity analysis, sentence-level novelty metrics, and longitudinal response comparison across dialogue turns.

Potential Limitations & Biases

Responses are generated by a single model (gpt-5-nano) and may reflect model-specific biases.
The dataset is synthetic and does not capture real human conversational variability.
Initial prompts are limited in number and may not span all possible domains or difficulty levels.
Response length is shaped by the 10-sentence system prompt, so lengths do not reflect the model's default verbosity. The limit is an instruction to the model rather than a hard truncation, so it is not strictly guaranteed for every response.
The dataset is static and reflects model behavior as of October 2026.

Contact

For questions, feedback, or collaboration inquiries related to this dataset, please contact:

jneb2004@byu.edu