---
pretty_name: ShareGPT Workload Cases
language:
- en
license: other
task_categories:
- text-generation
tags:
- synthetic
- conversations
- scheduling
- benchmarking
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - manifest.json
    - cases/**/*
---

# ShareGPT Workload Cases

This dataset packages workload-generation artifacts derived from `Aeala/ShareGPT_Vicuna_unfiltered`.
It includes the manifest describing how cases were generated and per-category case files intended for scheduler and systems benchmarking.

## Contents

- `manifest.json`: dataset generation metadata and category definitions.
- `cases/`: 176 generated case files across 22 workload categories.

## Generation Metadata

- Split: `train`
- Cases per category: `8`
- Tokenizer: `gpt-4` via `tiktoken`
- Generation seed: `7`

## Category Preview

- `steady_short`: Steady Poisson arrivals dominated by short prompts. (32 requests per case)
- `steady_mixed`: Steady mixed load with balanced request lengths. (64 requests per case)
- `bursty_chat`: Bursty conversational traffic with mostly short and medium prompts. (96 requests per case)
- `bursty_long`: Burst-heavy load with a larger share of long prompts. (72 requests per case)
- `flash_crowd`: Large clustered traffic spikes for stress testing scheduler fairness and queueing. (192 requests per case)
- `uniform_workload`: Uniform request sizes and steady arrivals for baseline throughput and scheduler overhead. (64 requests per case)
- `low_load`: Low-arrival workload with little contention to approximate ideal latency. (32 requests per case)
- `short_long_mix`: Majority short requests with a minority of long requests to expose head-of-line blocking. (80 requests per case)
- ...and 14 more categories.

## Notes

- Each category folder also contains a `category.json` file with category-level metadata.
- Source records are intentionally excluded from this Hub upload.
- The uploaded layout mirrors the local directory structure for reproducibility.