PrefixBench JSONL Datasets

These datasets generate deterministic prompts for testing KV prefix caching behavior in LLM inference servers such as vLLM and SGLang. The prompts use controlled shared prefixes plus small unique suffixes so benchmark clients can compare cache reuse, latency, and throughput across server configurations.

Files

• shared_schema_1k.jsonl: Simple shared-prefix benchmark. Every request reuses the same extraction instruction, JSON schema, and few-shot examples of about 1k approximate tokens, followed by a short unique document and query.
• same_document_multi_query_8k.jsonl: High-reuse long-document benchmark. Each document group reuses the same instruction, schema, and long source document of about 8k approximate tokens while changing only the final extraction question.
• multiturn_agent_branching.jsonl: Agent conversation benchmark. Prompts share a common system/tool root and then branch into simulated tool-use conversations, producing radix-tree-style prefix reuse across branch depths.
• eviction_pressure_4k.jsonl: Cache pressure benchmark. Many distinct prefix groups of about 4k approximate tokens are introduced first, then revisited later with new suffixes to test eviction and recovery behavior.

JSONL Schema

Each line is a standalone JSON object:

{
  "id": "unique_id",
  "prompt": "plain text prompt",
  "max_tokens": 128,
  "temperature": 0.0,
  "metadata": {
    "scenario": "scenario_name",
    "prefix_group": "shared_prefix_identifier",
    "expected_shared_prefix_tokens": 1000,
    "unique_suffix_tokens": 128,
    "request_order": 0,
    "description": "expected cache reuse behavior"
  }
}

Suggested Metrics

Track TTFT, TPOT, ITL, end-to-end latency, prefix cache hit rate, and throughput. Compare cold runs, warm runs, prefix caching disabled, and prefix caching enabled.