Traffic Crash Specialist
v1.0.0交通事故视频分析与检测专用技能。使用当需要：(1) 分析交通事故视频，(2) 事故识别、因果推理、预防分析，(3) 交通场景时空理解与对象定位，(4) 查询交通事故检测相关模型/数据集/论文，(5) 使用 CrashChat 或 Traffix VideoQA 进行视频问答，(6) 训练或评估交通视频分析模型。触...
⭐ 1· 41·0 current·0 all-time
by@sxy799
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name, description and included docs all describe traffic-crash video analysis, model training, evaluation and dataset references; the commands (git clone, conda env, pip install, huggingface download, training/eval scripts) are exactly what such a project would reasonably need.
Instruction Scope
SKILL.md focuses on cloning the CrashChat repo, setting up Python/torch, downloading model weights/datasets from GitHub/HuggingFace, and running training/evaluation scripts. It does not instruct reading unrelated system files, accessing unrelated environment variables, or exfiltrating data to unexpected endpoints.
Install Mechanism
This is an instruction-only skill (no install spec). The instructions require installing binaries/packages via pip/conda and downloading model weights from GitHub/HuggingFace. That is normal for ML projects, but installing a provided wheel file (flash_attn-...whl) or any binary wheel from the repo is an operation that requires trusting the repository — review the wheel and repo before running.
Credentials
The skill declares no required env vars, credentials, or config paths. The runtime instructions reference network endpoints appropriate for the task (GitHub, HuggingFace, PyTorch download index). No unrelated credentials are requested.
Persistence & Privilege
Registry flags show no always:true or other elevated persistence. As an instruction-only skill it does not request to modify other skills or system-wide agent settings.
Assessment
This skill appears coherent for traffic-accident video analysis. Before running anything: (1) inspect the GitHub repo and any included wheel files (e.g., flash_attn-*.whl) — running arbitrary wheels requires trust; (2) be prepared for large disk, GPU and bandwidth usage when downloading datasets/model weights and training; (3) verify dataset licensing and privacy (some sources like Nexar may contain personal data); (4) if a HuggingFace repo is private you may need credentials—do not supply unrelated secrets; (5) if you plan to let an agent run this autonomously, be aware it will perform network downloads and local installs as shown in SKILL.md.Like a lobster shell, security has layers — review code before you run it.
latestvk971fy0v60cav2kr71pxf2s17h83zz52
License
MIT-0
Free to use, modify, and redistribute. No attribution required.