isaacchung/controlled-generated-convos-gpt-4.1-mini

Controlled Generated Conversations: gpt-4.1-mini

Dataset Description

This dataset contains synthetic customer support conversations generated using gpt-4.1-mini as part of research on cross-lingual stability of LLM judges. The conversations are designed for evaluating how well language models maintain consistent performance across different languages, with a focus on Finno-Ugric languages (Estonian, Finnish, Hungarian) and English.

Dataset Summary

• Languages: Hungarian, Estonian, Finnish
• Domain: Customer support conversations
• Generator Model: gpt-4.1-mini
• Total Messages: 259,985
• Total Conversations: 30,000
• Generation Method: Controlled synthetic generation with structured parameters

Supported Tasks

• Cross-lingual evaluation: Compare model performance across languages
• Conversation analysis: Study dialogue patterns and structures
• Quality assessment: Evaluate linguistic quality (grammar, fluency, coherence)
• Label recovery: Classify conversation characteristics (industry, problem type, etc.)
• Ranking stability: Analyze consistency of model rankings across languages

Languages

Hungarian (hu)

• Messages: 79,928
• Conversations: 10,000
• Avg. conversation length: 8.0 messages
• Conversation length range: 1-18 messages
• Industries covered: 46
• Problem types: 20

Estonian (et)

• Messages: 100,800
• Conversations: 10,000
• Avg. conversation length: 10.1 messages
• Conversation length range: 4-18 messages
• Industries covered: 46
• Problem types: 20

Finnish (fi)

• Messages: 79,257
• Conversations: 10,000
• Avg. conversation length: 7.9 messages
• Conversation length range: 4-18 messages
• Industries covered: 46
• Problem types: 20

Dataset Structure

Data Instances

Each row represents a single message within a conversation, allowing for granular analysis of dialogue turns:

{
  "conversation_id": "gpt-4.1-mini_et_20260127-173115_0000",
  "message_id": 0,
  "from_name": "Klient",
  "from_type": "customer",
  "message": "Tere! Minu tellimus pidi juba kohal olema...",
  "subject": "Saatmise hilinemine tellimusele",
  "model": "gpt-4.1-mini",
  "language": "et",
  "industry": "Music & Audio",
  "problem": "Shipping delay",
  "channel": "chat",
  "agent_experience": "senior",
  "agent_type": "bot"
}

Data Fields

Core Message Fields

• conversation_id (string): Unique identifier for the conversation
• message_id (int): Position of message within conversation (0-indexed)
• from_name (string): Speaker name (e.g., "Customer", "Agent 1")
• from_type (string): Speaker role ("customer" or "agent")
• message (string): The actual message content

Conversation Metadata

• subject (string): Conversation topic/title
• model (string): Generator model used
• language (string): Language code (et, fi, hu, en-us)

Generation Parameters

• industry (string): Business domain (40+ categories)
• problem (string): Issue type (20+ categories)
• channel (string): Communication channel (email, chat)
• agent_experience (string): Agent level (junior, senior)
• agent_type (string): Agent type (human, bot)
• n_messages (int): Total messages in conversation
• n_agents (int): Number of agents involved

File Provenance

• file_model (string): Model from source filename
• file_language (string): Language from source filename
• file_datetime (string): Generation timestamp from filename

Data Splits

This dataset uses a single train split containing all conversations. For evaluation purposes, users can create their own splits based on:

• Language (cross-lingual evaluation)
• Industry/problem type (domain adaptation)
• Conversation length (complexity analysis)

Dataset Creation

Curation Rationale

This dataset was created to study cross-lingual stability of language models in evaluation tasks. Finno-Ugric languages (Estonian, Finnish, Hungarian) were specifically chosen as they are:

• Morphologically complex with rich inflectional systems
• Less represented in training data compared to English
• Suitable for testing model robustness across linguistic diversity

Source Data

Data Collection

Conversations were synthetically generated using structured prompts with controlled parameters:

1. Industry Selection: Random sampling from 40+ business domains
2. Problem Generation: 20+ customer support issue types
3. Channel & Agent: Varied communication contexts
4. Language: Native speaker patterns for each target language

Data Processing

• Message-level structuring: Each conversation split into individual messages
• Unique ID generation: Traceable identifiers linking messages to conversations
• Metadata preservation: All generation parameters maintained for analysis
• Quality validation: Automatic filtering for completeness and format

Annotation Process

No human annotation was performed. All labels represent the controlled generation parameters used to create each conversation, enabling evaluation of how well models can recover the original intent.

Considerations for Using the Data

Social Impact of Dataset

This dataset enables research into:

• Language equity: Understanding model performance disparities across languages
• Evaluation fairness: Developing more inclusive assessment methods
• Multilingual capabilities: Improving cross-lingual transfer in NLP systems

Discussion of Biases

Limitations:

• Synthetic nature: May not capture all nuances of real customer support
• Generation model bias: Inherits biases from the generator model (gpt-4.1-mini)
• Language representation: Quality may vary based on model's training data per language
• Cultural context: May not reflect local business practices or communication styles

Mitigation strategies:

• Use alongside real conversation data when possible
• Consider cultural and linguistic context in analysis
• Validate findings across multiple generator models

Other Known Limitations

• Temporal coverage: Generated at a single point in time
• Domain scope: Limited to customer support scenarios
• Generator dependency: Results tied to specific model capabilities
• Language variety: Standard varieties only, no dialectal variation

Additional Information

Dataset Curators

Created by the Cross-lingual Stability of LLM Judges research project.

Licensing Information

Licensed under MIT License. See LICENSE file for details.

Citation Information

If you use this dataset in your research, please cite:

@misc{chung2026crosslingualstabilityllmjudges,
      title={Cross-Lingual Stability of LLM Judges Under Controlled Generation: Evidence from Finno-Ugric Languages}, 
      author={Isaac Chung and Linda Freienthal},
      year={2026},
      eprint={2602.02287},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.02287}, 
}

Contributions

Thanks to the contributors of this dataset and the broader research community working on cross-lingual NLP evaluation.

Statistics

Dataset Overview

Metric	Value
Total Languages	3
Total Messages	259,985
Total Conversations	30,000
Average Messages per Conversation	8.7

Per-Language Statistics

Language	Messages	Conversations	Avg. Length
Hungarian (hu)	79,928	10,000	8.0
Estonian (et)	100,800	10,000	10.1
Finnish (fi)	79,257	10,000	7.9

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load specific language subset
dataset = load_dataset("isaacchung/controlled-generated-convos-gpt-4.1-mini", "et")
print(f"Estonian subset: {len(dataset)} messages")

# Load all subsets
dataset = load_dataset("isaacchung/controlled-generated-convos-gpt-4.1-mini")
print(f"Available languages: {list(dataset.keys())}")

# Access conversation metadata
et_data = dataset["et"]
first_msg = et_data[0]
print(f"Industry: {first_msg['industry']}")
print(f"Problem: {first_msg['problem']}")

Basic Analysis

# Group messages by conversation
conversations = {}
for msg in dataset["et"]:
    conv_id = msg["conversation_id"]
    if conv_id not in conversations:
        conversations[conv_id] = []
    conversations[conv_id].append(msg)

# Analyze conversation lengths
lengths = [len(conv) for conv in conversations.values()]
print(f"Avg conversation length: {sum(lengths)/len(lengths):.1f} messages")

Cross-lingual Analysis

# Compare industry distribution across languages
for lang in dataset.keys():
    industries = [msg["industry"] for msg in dataset[lang]]
    unique_industries = set(industries)
    print(f"{lang}: {len(unique_industries)} unique industries")