Long-context prompts in chat-message format at six bucket sizes
(1K, 2K, 4K, 8K, 16K, 32K user-message tokens), designed for benchmarking
LLM inference-server throughput across context-length regimes.
Each prompt is a 2-turn conversation (system + user) ready to send to any
OpenAI-compatible /v1/chat/completions endpoint. The system message
instructs the model to summarise the passage in exactly five words, so the
generated output is bounded… See the full description on the dataset page:
https://huggingface.co/datasets/nnilayy/pg19-concurrency-bench.
