sfc-gh-goliaro/kb-nano-balanced

kb-nano `balanced`工作负载是一个预计算的大型语言模型（LLM）吞吐量基准测试工作负载数据集，用于`kb_nano`项目的`tests/bench_vllm.py`测试。数据集来源于`allenai/WildChat-1M`，使用`meta-llama/Llama-3.1-8B-Instruct`作为参考分词器进行过滤和统计。数据集包含1000个请求，每个请求的提示令牌数在256到1024之间，解码令牌数上限为1024，下限为256。数据集以原始文本形式存储，允许基准测试运行器使用任何模型的分词器重新分词。数据集还提供了详细的令牌长度统计信息和模式描述。

The kb-nano `balanced` workload is a precomputed LLM throughput-benchmark workload used by `kb_nano`s `tests/bench_vllm.py`. The source dataset is `allenai/WildChat-1M`, and the reference tokenizer used for filtering and stats is `meta-llama/Llama-3.1-8B-Instruct`. The dataset contains 1000 requests, with prompt tokens ranging from 256 to 1024 and decode tokens capped at 1024 with a floor of 256. The chat content is stored as raw text, allowing the benchmark runner to re-tokenize the workload with any models tokenizer. Detailed token-length statistics and schema descriptions are also provided.