Dataset Card for Natural Conversation Benchmark (NC-Bench)

Dataset Summary

The Natural Conversation Benchmarks (NC-Bench) aim to answer the question: How well can generative AI converse like humans do? In other words, the benchmarks begin to measure the general conversational competence of large language models (LLMs). They do this by testing models' ability to generate an appropriate type of conversational action, or dialogue act, in response to a particular sequence of actions. The sequences of conversational actions, or patterns, are adapted from conversation science, specifically the model of sequence organization in the field of conversation analysis (Schegloff, 2007) and the pattern library of IBM Natural Conversation Framework (book). Models are tested by generating the next line in a transcript. NC-Bench is a lightweight method that is easily extensible to more conversation patterns.

The dataset consists of 720 samples:

• Basic (180): Patterns that capture basic practices of sequence management: answering inquiries, repairing answers, and closing pair sequences. This set uses ordinary conversational use cases and does NOT include passages for retrieval augmented generation (RAG). (See results above).
• RAG (180): Sequence management patterns from Basic but with the inclusion of a passage of information for RAG. Determining the faithfulness of the models' responses to the passage are not a primary goal. Instead, the goal is to determine if the model can maintain the conversation pattern in the face of a document context, which contains a competing language style and format. The uses cases in this set involve information giving using Wikipedia as a source.
• Complex Request (360): Sequence management patterns involving complex requests. Such requests require the agent to elicit details from the fictional user (for example, slot filling). Other patterns involve preliminaries to the inquiry-answer pair (i.e., pre-expansions). These use cases are business related.

Set	Pattern	Count	Pattern	Count
Basic	Inquiry	20	Paraphrase Request	20
	Incremental Request	20	Repeat Request	20
	Self-Correction	20	Example Request	20
	Definition Request	20	Sequence Closer	20
	Sequence Abort	20	Overview	180
RAG	Inquiry	20	Paraphrase Request	20
	Inquiry (Ungrounded)	20	Repeat Request	20
	Incremental/Correction	20	Example Request	20
	Definition Request	20	Sequence Closer	20
	Sequence Abort	20	Overview	180
Complex Request	Preliminary	40	Paraphrase Request	20
	Recommendation	60	Repeat Request	40
	Detail Request	60	Example Request	20
	Expansion	40	Sequence Closer	20
	Self-Correction	20	Sequence Abort	20
	Definition Request	20	Overview	360

Intended Uses

• Benchmarking: An expert-annotated benchmark for evaluating the general conversational competence of large language models (LLMs)

Limitations

• Size: With <1K samples, the dataset is best suited for evaluation, not large-scale training.

Leaderboard

Evaluation Code & Criteria.

The full evaluation pipeline, judge setup, and per-certierion scoring logic, is available on GitHub.

• Qwen-3B achieves the highest conversation competence on the Basic set with 82.22%.
• Granite-8B performs best on the RAG set with 77.77%.
• Granite-2B performs best on the Complex Request set with 80.15% accuracy.

Model	Basic (%)	RAG (%)	Complex Request (%)
granite-2B	72.22	76.11	80.15
granite-8B	76.11	77.77	77.04
Llama-3B	66.66	60.00	67.80
Llama-8B	68.88	68.88	71.06
Qwen-3B	82.22	75.55	62.19
Qwen-7B	80.55	73.88	76.06

Note: Model abbreviations:

• granite-2B = granite-3.3-2b-inst
• granite-8B = granite-3.3-8b-inst
• Llama-3B = Llama-3.2-3B-Inst
• Llama-8B = Llama-3.1-8B-Inst
• Qwen-3B = Qwen2.5-3B-Inst
• Qwen-7B = Qwen2.5-7B-Inst

Reference

Emanuel A. Schegloff, Sequence Organization in Interaction: A Primer in Conversation Analysis (Cambridge University Press, 2007)

How to cite this work

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

BibTeX

@article{moore2026ncbench,
      title={NC-Bench: An LLM Benchmark for Evaluating Conversational Competence}, 
      author={Robert J. Moore and Sungeun An and Farhan Ahmed and Jay Pankaj Gala},
      journal={arXiv preprint arXiv:2601.06426},
      year={2026},
      eprint={2601.06426},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.06426}, 
}