Giri2241/customer-support-client-agent-conversations

Customer Support Client-Agent Conversations Dataset

A synthetic context-summarized multi-turn customer-service question-answering dataset for banking domain conversations, designed for training and evaluating small language models on dialogue continuity and contextual understanding tasks.

Dataset Description

This dataset contains 183,337 context-summarized multi-turn customer-service conversations spanning various banking scenarios including account management, transactions, card services, fraud prevention, and general inquiries. Each instance includes conversation history summarization, client questions, and refined agent responses optimized for training instruction-tuned language models.

Domain: Banking Customer Service
Language: English
Task: Context-Summarized Multi-Turn Question Answering
Source: Derived from TalkMap Banking Conversation Corpus
Processing: Synthetic data construction with GPT-4o-mini summarization and GPT-4.1 response refinement

Dataset Statistics

Split	Samples	Total Turns	Avg Turns	Min Turns	Max Turns	Total Tokens	Avg Tokens
Train	128,335	1,291,138	10.06	2	53	37,478,648	292.04
Validation	18,333	183,364	10.00	2	50	5,348,170	291.72
Test	36,669	368,650	10.05	2	58	10,714,768	292.20
Total	183,337	1,843,152	10.05	2	58	53,541,586	292.02

Token counts computed using GPT-4 tokenizer

Dataset Construction Pipeline

The dataset was created through a controlled multi-stage processing pipeline:

1. Initial Data Source

• Base Corpus: TalkMap Banking Conversation Corpus (301,822 conversations)
• Original Format: Single-turn QA pairs in sequential order

2. Preprocessing & Filtering

• Retained conversations with 5-100 turns (realistic dialogue depth)
• Removed extremely short (<5 turns) and anomalously long (>100 turns) interactions
• Applied regex-based noise removal for formatting artifacts
• Result: ~200,000 filtered conversations

3. Multi-Turn Conversation Construction

• Aggregated sequential single-turns into structured multi-turn dialogues
• Applied de-duplication to remove redundant conversations
• Random partitioning into conversation stages:
  • Early stage: 20% (issue identification)
  • Mid stage: 70% (core interaction and information exchange)
  • Late stage: 10% (resolution and closure)

4. Context Summarization

• Model: GPT-4o-mini (max_tokens: 250, temperature: 0.3)
• Strategy: History summarization preserving essential conversational context
• Includes:
  • Client's primary issue/request and current status
  • Client and agent names (when mentioned)
  • Verification steps completed or pending
  • Exact account identifiers, dates, amounts, actions
  • Commitments, deadlines, scheduled follow-ups
  • Current conversation status

5. Response Refinement

• Model: GPT-4.1 (temperature: 0.4)
• Improvements:
  • Natural, human-like speaking patterns
  • Appropriate response length for question complexity
  • Clarity and precision
  • Contextual coherence with conversation history
  • Noise removal from original agent responses

6. Content Moderation

• OpenAI Moderation API applied to filter potentially offensive content
• Final quality assurance and validation

Data Fields

Each instance in the dataset contains the following fields:

• instruction (string): System-level instruction for the customer service agent
• history (string): Full conversation history (all prior turns)
• history_summary (string): GPT-4o-mini generated summary of conversation history
• client_question (string): Current client question/request
• agent_answer (string): Original agent response
• refined_agent_answer (string): GPT-4.1 refined agent response (used as ground truth)
• conversation_id (string): Unique identifier for the conversation

Intended Uses

Primary Use Cases

• Training small language models (SLMs) for customer-service QA
• Evaluating dialogue continuity and contextual understanding
• Research on context summarization strategies
• Multi-turn conversation modeling
• Privacy-preserving customer service automation

Supported Tasks

• Multi-turn question answering
• Dialogue state tracking
• Context-aware response generation

Usage Example

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("Lakshan2003/customer-support-client-agent-conversations")

# Load specific splits
train_dataset = load_dataset("Lakshan2003/customer-support-client-agent-conversations", split="train")
val_dataset = load_dataset("Lakshan2003/customer-support-client-agent-conversations", split="validation")
test_dataset = load_dataset("Lakshan2003/customer-support-client-agent-conversations", split="test")

# Access a sample
sample = train_dataset[0]
print("Instruction:", sample['instruction'])
print("History Summary:", sample['history_summary'])
print("Client Question:", sample['client_question'])
print("Agent Answer:", sample['refined_agent_answer'])

Training Example (QLoRA Fine-tuning)

from datasets import load_dataset
from transformers import AutoTokenizer

# Load dataset
dataset = load_dataset("Lakshan2003/customer-support-client-agent-conversations")

# Prompt template (example for LLaMA models)
prompt_template = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{instruction}<|eot_id|>

<|start_header_id|>user<|end_header_id|>

Conversation History:
{history_summary}

Client Question:
{client_question}

<|eot_id|>

<|start_header_id|>assistant<|end_header_id|>
{refined_agent_answer}<|eot_id|>"""

# Format function
def format_sample(sample):
    return prompt_template.format(
        instruction=sample['instruction'],
        history_summary=sample['history_summary'],
        client_question=sample['client_question'],
        refined_agent_answer=sample['refined_agent_answer']
    )

# Tokenize and prepare for training
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
tokenized_dataset = dataset.map(
    lambda x: tokenizer(format_sample(x), truncation=True, max_length=512),
    batched=True
)

Conversation Stage Distribution

The dataset includes conversation stage annotations for stage-based evaluation:

• Early Stage (~20%): Initial issue identification, greeting, authentication
• Mid Stage (~70%): Core information exchange, problem diagnosis, solution discussion
• Late Stage (~10%): Resolution confirmation, closure, follow-up scheduling

This distribution enables stage-wise performance analysis of models across different phases of customer-service interactions.

Limitations

• Synthetic Data: Generated from base corpus with LLM-based processing; may not capture full real-world variability
• Domain-Specific: Focused on banking domain; may require adaptation for other sectors
• Language: English only
• Privacy: Original sensitive information replaced/anonymized during synthesis
• Response Quality: Dependent on GPT-4.1 refinement; may reflect certain stylistic biases
• Turn Distribution: Filtered to 5-100 turns; excludes very short and extremely long conversations

Ethical Considerations

• Privacy Preservation: Synthetic data construction ensures no real customer PII is exposed
• Bias Mitigation: Response refinement aimed to reduce noise, but may introduce LLM-specific biases
• Intended Use: Designed for research and development; production deployment requires human oversight
• Responsible AI: Models trained on this data should be evaluated for fairness, safety, and accuracy before deployment

Citation

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

@article{cooray2026small,
  title={Can Small Language Models Handle Context-Summarized Multi-Turn Customer-Service QA? A Synthetic Data-Driven Comparative Evaluation},
  author={Cooray, Lakshan and Sumanathilaka, Deshan and Raju, Pattigadapa Venkatesh},
  journal={arXiv preprint arXiv:2602.00665},
  year={2026}
}

Dataset Curators

Primary Curator: Lakshan Cooray
Institution: Informatics Institute of Technology, Colombo, Sri Lanka
Contact: lakshan.20221470@iit.ac.lk
Co-authors: Deshan Sumanathilaka (Swansea University), Pattigadapa Venkatesh Raju (Zame AI)

License

This dataset is derived from the TalkMap Banking Conversation Corpus. Please refer to the original dataset's license for usage terms. The synthetic processing and enhancements are provided for research and educational purposes.

Related Resources

Fine-tuned Models

The following models have been trained on this dataset and are available on Hugging Face:

8B Models:

• Lakshan2003/Llama-3.1-8B-Instruct-customerservice
• Lakshan2003/Qwen-3-8B-Instruct-customerservice

3-4B Models:

• Lakshan2003/Qwen-3-4B-Instruct-customerservice
• Lakshan2003/Llama3.2-3B-instruct-customerservice
• Lakshan2003/Phi-4-mini-instruct-customerservice
• Lakshan2003/Gemma3-4B-instruct-customerservice
• Lakshan2003/SmolLM3-3B-instruct-customerservice

1-2B Models:

• Lakshan2003/Qwen-3-1.7B-instruct-customerservice
• Lakshan2003/Llama3.2-1B-instruct-customerservice

Paper

• arXiv: 2602.00665
• Title: Can Small Language Models Handle Context-Summarized Multi-Turn Customer-Service QA?

Acknowledgments

We thank:

• TalkMap for the original Banking Conversation Corpus
• OpenAI for GPT-4o-mini and GPT-4.1 API access
• Zame AI for funding support for API usage
• Informatics Institute of Technology for research infrastructure
• Human evaluators who contributed to qualitative assessment

Version History

• v1.0.0 (2026-02): Initial release with 183,337 conversations across train/val/test splits

Contact

For questions, issues, or collaboration opportunities:

• Email: lakshan.20221470@iit.ac.lk
• Institution: Informatics Institute of Technology, Colombo, Sri Lanka
• GitHub Issues: Please report any dataset issues via the Hugging Face dataset page