build-small-hackathon/limp-mode-traces
收藏资源简介:
该数据集包含来自Limp Mode的完整评估轨迹,Limp Mode是一个离线路边副驾驶系统。每个记录对应202个案例黄金套件中的一次诊断,涵盖驾驶员输入、检索器返回的内容、模型回答、触发的确定性安全规则以及安全底线后的最终裁决。数据集包含774条轨迹,跨越多个运行版本,如base-202-cleanpipe、baseline-202-pretraining、fixed-floor-pipeline、leap1-ft-202、tuned-rag-pipeline、zeroshot-full-pipeline和zeroshot-model-alone。关键列包括:model_verdict与final_verdict对比显示所有安全底线干预;dangerous_as_safe指标必须为0(表示预期为STOP但回答为DRIVE的情况);knowledge_surfaced跟踪隐藏知识(如惯性开关、电动汽车12V电池等)是否影响答案。这些轨迹记录了项目的主要发现:朴素RAG导致准确率从88%降至59%而知识覆盖率上升、危险标志底线升级错误以及微调前后的变化。
This dataset contains full evaluation trajectories from Limp Mode, an offline roadside co-pilot system. Each record corresponds to a diagnostic run within the gold-standard test suite of 202 cases, covering driver inputs, retrieved content from the retriever, model responses, triggered deterministic safety rules, and the final adjudication post-safety threshold enforcement. The dataset includes 774 trajectories spanning multiple operational versions, including base-202-cleanpipe, baseline-202-pretraining, fixed-floor-pipeline, leap1-ft-202, tuned-rag-pipeline, zeroshot-full-pipeline, and zeroshot-model-alone. Key columns include: the comparison between `model_verdict` and `final_verdict` to reveal all safety threshold interventions; the `dangerous_as_safe` metric which must equal 0 (referring to cases where the expected output was STOP but the model response was DRIVE); and `knowledge_surfaced`, which tracks whether hidden knowledge (e.g., inertia switches, 12V batteries for electric vehicles, etc.) impacts the model's outputs. These trajectories document the project's core findings: naive RAG reduces model accuracy from 88% to 59% while increasing knowledge coverage, errors in safety threshold escalation for hazardous scenarios, and performance changes observed before and after fine-tuning.




