---
language:
- en
task_categories:
- question-answering
- summarization
- text-generation
pretty_name: LoopServe Multi-Turn Dialogue Benchmark
tags:
- llm
- kv_cache
configs:
- config_name: conversations
  data_files: conversations.jsonl
- config_name: multi_turn_few_shot_learning
  data_files: multi_turn/few_shot_learning/*.jsonl
- config_name: multi_turn_needle_in_haystack
  data_files: multi_turn/needle_in_haystack/*.jsonl
- config_name: multi_turn_question_answering
  data_files: multi_turn/question_answering/*.jsonl
- config_name: multi_turn_summarization
  data_files: multi_turn/summarization/*.jsonl
- config_name: single_turn_few_shot_learning
  data_files: single_turn/few_shot_learning/*.jsonl
- config_name: single_turn_needle_in_haystack
  data_files: single_turn/needle_in_haystack/*.jsonl
- config_name: single_turn_question_answering
  data_files: single_turn/question_answering/*.jsonl
- config_name: single_turn_summarization
  data_files: single_turn/summarization/*.jsonl
---

This repository contains the benchmark datasets proposed in the paper **[LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues](https://huggingface.co/papers/2507.13681)**.

The LoopServe benchmark introduces eleven multi-turn datasets designed to evaluate large language models (LLMs) on realistic query positions and conversational dependencies. This is crucial for assessing LLM inference acceleration methods in dynamic, multi-turn dialogue settings common in applications like chatbots and virtual assistants.

**Paper:** [LoopServe: An Adaptive Dual-phase LLM Inference Acceleration System for Multi-Turn Dialogues](https://huggingface.co/papers/2507.13681)

### Sample Usage

You can load different subsets of the dataset using the `load_dataset` function from the `datasets` library. For example, to load the `multi_turn_question_answering` subset:

```python
from datasets import load_dataset

# Load the multi-turn question-answering subset
dataset_qa_multi = load_dataset("MKV_Cache", "multi_turn_question_answering")
print(dataset_qa_multi)

# Load the single-turn summarization subset
dataset_sum_single = load_dataset("MKV_Cache", "single_turn_summarization")
print(dataset_sum_single)

# Load the base conversations data
dataset_conv = load_dataset("MKV_Cache", "conversations")
print(dataset_conv)
```

### Dataset Structure

The repository contains the following file structure for the benchmark data:

``` shell
.
├── README.md
├── conversations.jsonl
├── multi_turn
│   ├── few_shot_learning
│   ├── needle_in_haystack
│   ├── question_answering
│   └── summarization
└── single_turn
    ├── few_shot_learning
    ├── needle_in_haystack
    ├── question_answering
    └── summarization
```