dataset-A-routing
收藏资源简介:
该数据集名为A - Routing (3 modelli, verdict-level),是一个用于评估大语言模型(LLM)路由系统的基准数据集。它专注于六个核心能力维度:规划与代理、数学推理、编程、指令遵循、世界知识和创造性合成。数据集包含5504个查询,每个查询均在三个预训练模型(qwen3.5-9b, deepseek-v4-flash, kimi2.6)上执行,并采用混合评估协议:对于数学、编程和指令遵循任务使用确定性评分器(如数学等价性检查、单元测试通过率、ifeval约束检查),对于规划与代理任务使用LLM评委面板(2-of-3多数表决),对于创造性合成和部分世界知识任务使用单LLM评委。数据集提供三种配置:1) results配置:提供核心的查询级判断结果,包含查询ID、查询文本(受限制的查询会被掩码)、能力维度、评估协议、数据来源以及三个模型各自的正确性布尔值(可能为空值表示弃权)。2) verbose配置:在results基础上,为每个模型额外提供原始响应、思维链、推理成本、延迟、令牌使用情况、完成原因、评分器元数据等详细信息。3) evals配置:仅包含提示和真实值,用于复现和重新运行评估测试。数据规模在1千到1万样本之间。该数据集旨在支持LLM路由策略的评估、模型能力比较以及多领域任务下的性能分析。数据集基于CC-BY-4.0许可证发布,但源自其他受限制数据集的查询文本仍遵循其原始许可证。
This dataset, named A-Routing (3 modelli, verdict-level), is a benchmark dataset for evaluating Large Language Model (LLM) routing systems. It focuses on six core capability dimensions: Planning & Agent, Mathematical Reasoning, Programming, Instruction Following, World Knowledge, and Creative Synthesis. The dataset contains 5,504 queries, each executed on three pre-trained models: qwen3.5-9b, deepseek-v4-flash, and kimi2.6. It adopts a hybrid evaluation protocol: deterministic scorers (such as mathematical equivalence checks, unit test pass rates, and ifeval constraint checks) are used for mathematical reasoning, programming, and instruction following tasks; an LLM judge panel (2-of-3 majority vote) is used for planning & agent tasks; and a single LLM judge is used for creative synthesis and partial world knowledge tasks. The dataset offers three configurations: 1) Results configuration: Provides core query-level judgment results, including query ID, query text (restricted queries will be masked), capability dimension, evaluation protocol, data source, and the correctness boolean values of each of the three models (empty values indicate abstentions). 2) Verbose configuration: Based on the Results configuration, it additionally provides detailed information for each model, including original response, Chain-of-Thought (CoT), inference cost, latency, token usage, completion reason, and scorer metadata. 3) Evals configuration: Only includes prompts and ground truth values, used for reproducing and re-running evaluation tests. The dataset scale ranges from 1,000 to 10,000 samples. This dataset aims to support the evaluation of LLM routing strategies, model capability comparison, and performance analysis across multi-domain tasks. The dataset is released under the CC-BY-4.0 license, but query texts originating from other restricted datasets still follow their original licenses.




