eigen-ai-labs/95-cache-dataset

Prefix Cache Benchmark Datasets

Datasets for benchmarking DeepSeek-V3 prefix cache performance with ~95% cache hit rate.

Dataset Variants

Dataset	Source	Samples	Avg Tokens	Cache Hit Rate
k8s_dataset.jsonl	Kubernetes docs	1,000	~70,000	~95%
sglang_dataset.jsonl	SGLang source code	1,000	~70,058	~95%
yahoo_finance_dataset.jsonl	SGLang + Yahoo Finance	1,000	~70,057	~95%
combined_dataset.jsonl	All three combined	3,000	~70,000	~95%

Structure

All datasets share:

• Shared prefix: 66,500 tokens (same for all requests) → cached
• Variable suffix: ~3,500 tokens (different per request) → not cached

{
  "messages": [
    {"role": "system", "content": "<66.5K token shared prefix>"},
    {"role": "user", "content": "<~3.5K token unique suffix + query>"}
  ],
  "metadata": {
    "shared_prefix_tokens": 66500,
    "total_input_tokens": 70057,
    "cache_hit_rate": 0.9492,
    "source": "sglang"
  }
}

Quick Start

Option 1: Benchmark Script (Recommended, Shuffled)

Auto-downloads dataset and shuffles samples each run for varied request ordering.

pip install evalscope huggingface_hub

# Download script
wget -q https://huggingface.co/datasets/eigen-ai-labs/95-cache-dataset/resolve/main/run_benchmark.sh
chmod +x run_benchmark.sh

# Run benchmark (auto-downloads dataset)
./run_benchmark.sh --url http://localhost:8000/v1/chat/completions --number 500 --parallel 4

Examples

# Quick test
./run_benchmark.sh --url http://localhost:8000/v1/chat/completions --number 100

# Specific dataset variant
./run_benchmark.sh --url http://localhost:8000/v1/chat/completions --variant k8s --number 500

# Reproducible run
./run_benchmark.sh --url http://localhost:8000/v1/chat/completions --number 500 --seed 42

Options

Option	Description	Default
--url	API endpoint (required)	-
--variant	Dataset: combined, k8s, sglang, yahoo_finance	combined
--number	Number of requests	500
--parallel	Concurrent requests	1
--max-tokens	Max output tokens	200
--seed	Random seed for shuffle	random
--model	Model name	deepseek-ai/DeepSeek-V3
--dataset-path	Custom dataset (overrides variant)	-

Option 2: EvalScope Directly (Sequential)

Runs samples in sequential order without shuffling. Supports multi-config runs.

DATASET_PATH=$(python -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('eigen-ai-labs/95-cache-dataset', 'data/combined_dataset_openqa.jsonl', repo_type='dataset'))")

evalscope perf \
    --model deepseek-ai/DeepSeek-V3 \
    --url http://localhost:8000/v1/chat/completions \
    --api openai \
    --dataset openqa \
    --dataset-path "$DATASET_PATH" \
    --max-tokens 200 \
    --parallel 1 2 4 \
    --number 500 500 500 \
    --stream

File Structure

data/
├── combined_dataset.jsonl              # Original messages format (3000 samples)
├── combined_dataset_openqa.jsonl       # EvalScope-compatible format
├── k8s_dataset.jsonl                   # Original messages format (1000 samples)
├── k8s_dataset_openqa.jsonl            # EvalScope-compatible format
├── sglang_dataset.jsonl                # Original messages format (1000 samples)
├── sglang_dataset_openqa.jsonl         # EvalScope-compatible format
├── yahoo_finance_dataset.jsonl         # Original messages format (1000 samples)
└── yahoo_finance_dataset_openqa.jsonl  # EvalScope-compatible format

run_benchmark.sh                        # Benchmark script with auto-shuffle

Data Sources

Dataset	Prefix Source	Suffix Source
k8s	Kubernetes official docs	Kubernetes docs
sglang	SGLang source code	SGLang source code
yahoo_finance	SGLang source code	Yahoo Finance data
combined	Mixed (all above)	Mixed (all above)

License

Apache 2.0