Vihindi-K/ConvoDrift_conversational_dataset

ConvoDrift Conversational Dataset

Citation

Preprint link: https://arxiv.org/abs/2610.02873

Dataset Description

ConvoDrift is an English multi-turn conversational dataset designed for studying tone drift and the direction of tone changes during iterative writing tasks. Each conversation contains six user prompts and six corresponding assistant responses. The conversations cover five communication genres: business emails, casual emails, LinkedIn posts, quotes and wishes, and tweets.

The records include conversation-level metadata, turn-level drift and direction labels, validation results, quality ratings, and optional refined labels and feedback.

Dataset Subsets

Subset	Communication genre	Conversations	Prompt-response pairs
business_email	Business email	2,997	17,982
casual_email	Casual email	3,246	19,476
linkedin_post	LinkedIn post	3,150	18,900
quotes_wishes	Quotes and wishes	3,340	20,040
tweet	Tweet	2,994	17,964
Total		15,727	94,362

All files are exposed as a train split to support the Hugging Face Data Viewer. The dataset does not prescribe a benchmark train, validation, or test partition.

Loading the Dataset

Install the Hugging Face datasets library and select the required subset:

from datasets import load_dataset

dataset = load_dataset(
    "Vihindi-K/ConvoDrift_conversational_dataset",
    "business_email",
)

Available subset names are:

business_email
casual_email
linkedin_post
quotes_wishes
tweet

To load every subset:

from datasets import get_dataset_config_names, load_dataset

repo_id = "Vihindi-K/ConvoDrift_conversational_dataset"
datasets_by_genre = {
    name: load_dataset(repo_id, name)
    for name in get_dataset_config_names(repo_id)
}

Replace the repository identifier in these examples if the published repository uses a different name.

Data Structure

Each JSONL line represents one conversation and follows this general structure:

{
  "status": "ok",
  "data": {
    "conversation_id": "conv_00586",
    "communication_genre": "business email",
    "scenario": "Client Meeting & Sales",
    "pairs": [
      {
        "prompt": "...",
        "response": "...",
        "drift": false,
        "direction": 0
      }
    ],
    "model": "gpt-4.1-mini-2025-04-14",
    "validation": {
      "valid_json": true,
      "six_pairs": true,
      "first_drift_false": true,
      "min_two_drifts": true,
      "is_valid": true
    }
  },
  "ratings": {
    "Q1": 3,
    "Q2": 3,
    "Q3": 5,
    "Q4": 3,
    "Q5": 5,
    "Q6": 5,
    "Q7": 5,
    "Q8": 5
  },
  "refined_drift_label": [null, null, false, true, null, null],
  "refined_direction_label": [null, null, "0", "2", null, null],
  "feedbacks": {
    "Q1": "...",
    "Q2": "...",
    "Q3": "",
    "Q4": "",
    "Q5": "",
    "Q6": "",
    "Q7": "",
    "Q8": ""
  }
}

The direction field may be absent when no direction was assigned to a turn.

Field Descriptions

Field	Description
status	Record-processing status.
data.conversation_id	Unique conversation identifier within the dataset.
data.communication_genre	Communication genre represented by the conversation.
data.scenario	Writing scenario or topical context.
data.pairs	Ordered list of six prompt-response pairs.
data.pairs[].prompt	User instruction for the current turn.
data.pairs[].response	Assistant response to the instruction.
data.pairs[].drift	Whether the turn requests a tone shift.
data.pairs[].direction	Direction of the requested tone shift, when available.
data.model	Model used to generate the conversation.
data.validation	Structural and label-related validation checks.
ratings.Q1-ratings.Q8	Quality ratings from the dataset evaluation protocol.
refined_drift_label	Optional corrected drift labels aligned with the six turns.
refined_direction_label	Optional corrected direction labels aligned with the six turns.
feedbacks.Q1-feedbacks.Q8	Optional evaluator feedback corresponding to each rating criterion.

Drift and Direction Labels

drift indicates whether a prompt requests a change in tone:

• false: no tone shift is requested.
• true: a tone shift is requested.

direction field indicates whether the response tone has actually shifted.

• 0: no tone shift.
• 1: more formal to more casual.
• 2: more casual to more formal.

The refined-label arrays are positionally aligned with data.pairs. A null value in a refined-label array does not mean that the turn has no label. It means that no correction was supplied, so the original drift or direction value from the corresponding prompt-response pair should be used.

For example, the effective label for turn i can be resolved as follows:

effective_drift = (
    row["refined_drift_label"][i]
    if row["refined_drift_label"][i] is not None
    else row["data"]["pairs"][i].get("drift")
)

effective_direction = (
    row["refined_direction_label"][i]
    if row["refined_direction_label"][i] is not None
    else row["data"]["pairs"][i].get("direction")
)

Intended Uses

The dataset can support research on:

• tone-shift and conversational-drift detection;
• direction classification for formal-to-casual and casual-to-formal changes;
• controlled text rewriting across communication genres;
• multi-turn instruction following and response consistency;
• evaluation of style-preserving language models.