---

name: llm-workflows description: "Guide DeepKE-LLM instruction KGC, OneKE, ICL/API, CodeKGC, and large-model workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

DeepKE LLM workflows

Use this sub-skill when a task asks for DeepKE-LLM, instruction-based knowledge graph construction, OneKE, InstructKGC, LLM in-context extraction, LLM-powered data augmentation, UnleashLLMRE, CodeKGC, CPM-Bee, or large-model fine-tuning/inference around information extraction.

Route by intent

• Choose among OneKE, InstructKGC, ICL/API, CodeKGC, UnleashLLMRE, or CPM-Bee: read references/workflows.md for the workflow decision table and resource expectations.
• Prepare instruction data or inspect JSONL schemas: read references/data-formats.md and use scripts/convert_ie_instruction.py for a safe, standalone conversion path for common NER/RE/SPO-style records.
• Check local large-model dependencies, API environment variables, CUDA visibility, and optional paths: run scripts/check_llm_workflow_env.py. It does not load models or call remote APIs.
• Debug GPU memory, LoRA/P-tuning, OpenAI-compatible API, JSONL, prompt, or converter failures: read references/troubleshooting.md.
• Compare model/runtime families: read references/model-and-runtime-overview.md before mixing dependencies from classic DeepKE and DeepKE-LLM.

What this sub-skill owns

• DeepKE-LLM instruction-based information extraction and knowledge graph construction.
• InstructKGC-style conversion, LoRA fine-tuning, P-tuning, and OpenDelta workflows for LLaMA/ZhiXi, ChatGLM, MOSS, Baichuan, Qwen, CPM-Bee, and related large models.
• OneKE schema-based bilingual information extraction planning and inference readiness.
• LLMICL-style OpenAI-compatible in-context information extraction, data augmentation, and CCKS instruction KGC.
• UnleashLLMRE few-shot relation-extraction prompting/data augmentation.
• CodeKGC code-style prompts and schema/in-context/test-example files for code LLMs.

What this sub-skill does not own

• Classic BERT/CRF/CNN/RNN/GCN supervised NER, RE, AE, EE, multimodal, document, few-shot, or cnSchema quick-load workflows; route those to supervised-extraction.
• PRGC, PURE, ASP, and MT5 non-LLM triple-extraction workflows; route those to triple-extraction.
• Generic weak supervision and distant supervision data prep for classic DeepKE examples; route those to data-preparation.
• Deploying the DeepKE MCP server/client wrapper; route that to mcp-tools.

Quick operating pattern

1. Decide whether the user wants data conversion, prompt/API inference, local model inference, LoRA/P-tuning/OpenDelta fine-tuning, evaluation, or troubleshooting.
2. Confirm resources before expensive work: GPU type/memory/count, model checkpoint availability, API credentials, dataset size, output directory, and network/cache policy.
3. Run python scripts/check_llm_workflow_env.py --workflow <workflow> to inspect packages and environment variables without loading the model.
4. For instruction data, inspect JSONL line-by-line. DeepKE-LLM tooling commonly writes JSONL, not one large JSON array.
5. Use the workflow reference to set task-specific fields: task, source, instruction, schema, input, output, model name/path, LoRA/output directories, API endpoint, and prompt templates.
6. Do not silently start model downloads, API calls, DeepSpeed jobs, or multi-GPU fine-tuning. Confirm cost and runtime first.

Safety and verification stance

The bundled scripts are safe diagnostics and data shapers. They do not train models, call OpenAI-compatible APIs, download model weights, run OneKE, or launch DeepSpeed. Large-model runtime success still requires user-provided checkpoints or API credentials and suitable hardware.