---

name: prompt-and-inference description: "Build and debug OpenCompass prompt templates, PromptList dialogue prompts, retrievers, Gen/PPL inferencers, reader_cfg answer masking, in-context examples, and meta_template role routing." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

OpenCompass Prompt and Inference

Use this sub-skill when an OpenCompass task involves prompt construction or inference config behavior: PromptTemplate, RawPromptTemplate, PromptList dialogue turns, ice_template, prompt_template, retrievers, GenInferencer, PPLInferencer, output-column masking, or meta_template role mapping.

Route the Task

• For prompt syntax, answer masking, missing literal fields, dialogue PromptList, raw messages, or CoT wording, use references/prompt-templates.md.
• For choosing ZeroRetriever/few-shot retrievers and wiring GenInferencer vs PPLInferencer, use references/retrievers-and-inferencers.md.
• For symptoms such as literal {answer}, leaked labels, missing system role handling, unsupported PPL candidate templates, or train split failures, use references/troubleshooting.md.
• For a no-download local preview of tiny samples and templates, run python scripts/render_prompt_preview.py --help from this sub-skill directory.

High-Value Defaults

• Use GenInferencer for free-form answers and keep the dataset answer column in reader_cfg.output_column so it is replaced by gen_field_replace_token instead of leaking into the generation prompt.
• Use PPLInferencer for multiple-choice/discriminative tasks and provide candidate templates keyed by label ("A", "B", etc.) in prompt_template.template or ice_template.template.
• Use ZeroRetriever for 0-shot; use FixKRetriever only when the selected in-context examples exist in the dataset train/index split.
• Put model-specific chat wrappers in the model meta_template; keep dataset prompt templates model-agnostic and role-based (HUMAN, BOT, SYSTEM).

Boundaries

• Route dataset class creation, reader_cfg source columns, and config file structure to configuration-and-datasets.
• Route tokenizer behavior, model chat templates, backend acceleration, and real HF/API inference issues to model-backends.
• Do not claim real model inference was verified from this sub-skill; the bundled script renders prompt construction only.