---

name: stanford-alpaca description: "Routes Stanford Alpaca instruction-data, Self-Instruct generation, supervised fine-tuning, and LLaMA weight-diff recovery workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Stanford Alpaca

Use this root skill to route research tasks involving Stanford Alpaca's released instruction-following data, its historical Self-Instruct-style generator, Hugging Face supervised fine-tuning flow, or Alpaca/LLaMA weight-diff recovery. It is a self-contained operational guide: use its bundled references and helpers rather than reopening a source checkout.

Start with the right route

User need or signal	Read or run
Alpaca JSON/JSONL schema, prompt text, label masking, datasheet/model-card questions, data validation	dataset-and-prompts
Seed tasks, prompt rendering, saved completion parsing, OpenAI completion API, regen.json, filtering, ROUGE-L deduplication	instruction-generation
Trainer, torchrun, FSDP, DeepSpeed, LLaMA/OPT SFT, OOM, batch-size planning, tokenizer/model preparation	fine-tuning
path_raw, path_diff, path_tuned, recovering or creating model weight differences, checksum, recovery inference	weight-diff-recovery

Safe starting sequence

1. For any new or generated corpus, run the offline validator in dataset-and-prompts first. It accepts JSON arrays or JSONL, checks the instruction/input/output contract, and can show prompt previews.
2. For instruction synthesis, debug the bundled prompt renderer and saved-response parser before using a live API. Live completion calls require credentials, network access, and a legacy completion-compatible OpenAI client.
3. For SFT, build a launch command before launching it. The bundled command builder only prints text; full LLaMA/OPT jobs require an appropriate model checkpoint, CUDA environment, and enough memory.
4. For weight recovery, use the bundled dry-run command builder to distinguish raw, diff, and output checkpoint paths before loading any tensors.

Runtime dependencies and capability gates

This is a script-oriented project rather than an installable Python distribution. To use every bundled helper, install a Python environment with PyTorch, Transformers, Fire, NumPy, rouge-score, sentencepiece, tokenizers, and an OpenAI client compatible with openai.Completion.create. A historically compatible baseline is transformers==4.28.1, tokenizers==0.13.3, openai==0.27.8, and numpy<2 with older PyTorch builds.

Run the bundled cross-skill probe after installation:

python scripts/check_stanford_alpaca_env.py

• Use --require-cuda only when a CUDA workflow is actually required; CPU imports do not validate multi-GPU FSDP, DeepSpeed, or large-checkpoint recovery.
• Add DeepSpeed only for the optional ZeRO-3 offload path. Add experiment tracking integrations only if you explicitly enable them in Transformers.
• Keep OpenAI API keys in the process environment, never in scripts, datasets, prompts, or skill files.

Boundaries and safety

• The released data and weight-diff artifacts carry non-commercial/research-use restrictions. Read dataset intended use and licenses before using or distributing derived data, models, or recovered weights.
• This skill provides documentation and safe planning helpers; it does not claim that a live OpenAI call, a full GPU fine-tune, or real LLaMA checkpoint recovery has run in a later user's environment.
• Do not treat the historical training hyperparameters as universally optimal. Validate data, model compatibility, hardware, and evaluation goals for the current experiment.

Shared references

• Cross-cutting troubleshooting: install/import compatibility, credentials, CUDA, data paths, artifacts, and licensing failures.
• Repository provenance: source commit and evidence baseline; read it before deciding whether a refresh is needed.
• Environment probe: inspect dependencies and optional CUDA availability without loading a model or calling a network service.

Common multi-step routes

• Generate then train: instruction-generation -> dataset-and-prompts -> fine-tuning.
• Use the released data directly: dataset-and-prompts -> fine-tuning.
• Recover then use a model: weight-diff-recovery -> fine-tuning only if additional supervised training is intended.