prnshv/TeleResilienceBench

TeleResilienceBench是一个电信领域的基准测试，用于评估模型的**推理延续韧性**：给定一个问题和一个部分完成（可能包含错误）的推理轨迹，模型必须继续推理并恢复正确的最终答案。该基准测试涵盖GSMA Open-Telco LLM套件中的七个电信子领域，实例是通过从一个弱生成器中获取失败的解决方案，在其错误轨迹的中点截断，并将该部分轨迹作为目标模型的延续上下文构建的。性能评估采用以恢复为中心的指标，主要包括**正确翻转率（CFR）**，以及**错误翻转率（WFR）**和**无翻转率（NFR）**。数据集包含两个主要文件：TeleResilienceBench.csv（主文件，多选题）和Auxiliary.csv（辅助文件，电信数学问题）。

`TeleResilienceBench` is a telecommunications benchmark for **reasoning continuation resilience**: given a question and a partially completed (and potentially flawed) reasoning trace, a model must continue the reasoning and recover the correct final answer. The benchmark spans seven telecom sub-domains from the GSMA Open-Telco LLM suite, and instances are constructed by taking failed solutions from a weak generator, truncating the flawed trace at its midpoint, and using that partial trace as the continuation context for target models. Performance is evaluated with recovery-focused metrics, primarily **Correct Flip Rate (CFR)**, alongside **Wrong Flip Rate (WFR)** and **No Flip Rate (NFR)**. The dataset includes two main files: `TeleResilienceBench.csv` (main, multiple-choice) and `Auxiliary.csv` (aux, telecom-math).