Step3-VL Finetune

Step3-VL-10B 多模态模型微调指南。用于在 GPU 服务器上进行 Step3-VL 模型的 LoRA/全量微调。包含配置、训练、推理完整流程。

This guide appears to be a legitimate GPU finetuning how-to, but take the following precautions before using it unchanged: - Treat the listed hostnames, container names, registry URLs, and internal IP (172.18.10.103) as environment-specific examples. Do not run commands that connect to those hosts unless you control/trust them. - Run any code (monkey patches and the custom save_adapter) in an isolated environment (dedicated GPU machine or container) and back up original model checkpoints first. The guidance deliberately bypasses PEFT checks and monkey-patches model internals — this can produce incompatible or unsafe artifacts if misapplied. - Verify the Docker image and any external services (harbor registry, vLLM endpoint) before pulling or sending data. Confirm licenses and data handling policies for the base model and any datasets used. - Confirm NCCL/CUDA environment settings match your cluster and drivers; incorrect NCCL tweaks can impact other jobs on shared nodes. - Inspect adapter_model.bin contents before sharing or uploading; the custom save routine produces a binary blob that could contain unexpected tensors. If you want a higher-confidence assessment, provide any code files (model_utils.py, dataset.py, inference.py) referenced in the guide or clarify whether the hostnames and endpoints are placeholders or part of a network the agent will reach — that would allow a more specific check for network/credential misuse.

Type: OpenClaw Skill Name: step3-vl-finetune Version: 1.0.0 The skill bundle contains sensitive internal infrastructure details, including a specific server address ([email protected]), an internal IP (172.18.10.103), and private registry paths (harbor.aibee.cn). The SKILL.md file also provides code for monkey-patching model forward passes and bypassing PEFT library checks to manually save weights. While these appear to be technical workarounds for a specific environment, the disclosure of internal network topology and the use of non-standard code execution patterns represent a significant information disclosure risk.