---

name: instruction-generation description: "Guide Self-Instruct-style instruction data generation with prompt rendering, completion parsing, filtering, ROUGE-L deduplication, and safe credential/network handling." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Instruction Generation

Use this sub-skill for the Stanford Alpaca instruction-generation workflow: render seed-task prompts, inspect or parse saved OpenAI completion text, apply the source filters, and understand how the live generator resumes through regen.json.

Use this sub-skill when

• You want to render prompts from seed records in the repo's prompt.txt style.
• You need to parse or debug saved completion text without calling OpenAI.
• You want to understand the live generation loop, filter rules, ROUGE-L deduplication, or resume behavior.
• You need a safe place to reason about credential, network, and legacy OpenAI client requirements.

Route elsewhere when

• The question is about released Alpaca dataset schema, license, or training-record formatting for SFT: see dataset-and-prompts.
• The question is about turning generated records into supervised fine-tuning input or Trainer data modules: see fine-tuning.
• The question is about checkpoint diffs or recovery: see weight-diff-recovery.

Bundled references

• references/workflows.md — offline prompt debugging, saved-completion parsing, live-generation gates, and regen.json resume semantics.
• references/prompt-template.md — bundled copy of the prompt template used by the generator.
• references/api-reference.md — verified signatures, defaults, and record schemas.
• references/troubleshooting.md — API key, rate limit, client compatibility, parsing, multiprocessing, and ROUGE/tokenization guidance.

Bundled scripts

• scripts/render_instruction_prompt.py — offline prompt renderer for tiny fixtures or user-provided seed-task files.
• scripts/parse_openai_response.py — offline parser for saved completion text or JSON payloads.

Safety and workflow notes

• Offline prompt rendering and saved-response parsing are deterministic and do not call OpenAI.
• Live generation requires OPENAI_API_KEY, network access, and a legacy OpenAI completion client compatible with openai.Completion.create.
• The generator loads regen.json from the output directory when present, appends accepted records, and rewrites that file after each batch.
• ROUGE-L deduplication uses the source thresholding behavior; if you are only debugging formatting, use the bundled scripts first and keep networked runs separate.

Quick orientation

1. Read the workflow notes and the API reference.
2. Render a tiny prompt fixture with the bundled renderer to validate formatting.
3. Parse a saved completion sample with the bundled parser to inspect the filter behavior.
4. Only move to live API generation after the credential and network gate is satisfied.