Format Adapter Alpaca

format-adapter-alpaca

Convert canonical training example records (@agentic/code/frameworks/training-complete/schemas/example-record.yaml) into Alpaca-format JSONL for downstream SFT training frameworks. Alpaca is the original Stanford self-instruct format and remains widely supported by trainers like Axolotl, LLaMA-Factory, and Unsloth.

When to Use

• Emitting a training split for a tuner that consumes Alpaca JSONL
• Shipping a public dataset in the lowest-common-denominator instruction format
• Interoperating with legacy pipelines that predate ChatML/ShareGPT

Parameters

• <input-glob> (required) — glob of canonical records (e.g., examples/raw/*.json)
• --output <path> (optional) — output JSONL path. Default: .aiwg/training/exports/alpaca-<timestamp>.jsonl
• --validate-round-trip (optional) — reload output and diff against canonical invariants before succeeding

Format Spec

One JSON object per line with fields {instruction, input, output}:

{"instruction": "You are a helpful assistant.", "input": "Explain photosynthesis in one sentence.", "output": "Photosynthesis is the process by which plants convert sunlight, water, and CO2 into glucose and oxygen."}

Operation

1. Load canonical records — parse each file per example-record.yaml; reject invalid records.
2. Transform — map fields:
   • instruction ← input.system (fallback to input.user if no system prompt)
   • input ← input.user (empty string "" if input.system was empty and input.user was promoted to instruction)
   • output ← output.assistant
3. Validate target — ensure instruction and output are non-empty; reject preference/tool_use records (not representable — route to sharegpt/chatml adapter).
4. Round-trip check (if --validate-round-trip) — parse output back and confirm canonical invariants (id, task_type, input.user, output.assistant, quality_grade, license, provenance_id) survive via sidecar.
5. Write output + log — emit JSONL, write sidecar, append format-convert event via memory-log-append.

Round-Trip Invariants

Alpaca fields cover only input.user and output.assistant. All other invariant fields (id, task_type, quality_grade, license, provenance_id) are preserved via sidecar.

Sidecar Metadata

Written alongside output as <output>.metadata.yaml — contains a list keyed by line number with: id, task_type, metadata.*, output.reasoning_trace, output.tool_calls, input.context_refs, and input.tools_available. Reasoning traces and tool calls are structural losses in Alpaca — always go to sidecar.

Acceptance Criteria

• Every canonical record produces exactly one Alpaca line (or is rejected with a logged reason).
• --validate-round-trip reconstructs canonical invariants 100% from (JSONL + sidecar).
• Preference and tool_use records are rejected with a pointer to the correct adapter.
• A format-convert event is logged with input count, output count, and rejection count.

References

• Alpaca methodology — Taori et al. 2023 (Stanford Alpaca, implied)
• ADR-022 D7 — canonical + adapter strategy

Delegation

• @agentic/code/addons/semantic-memory/skills/memory-log-append/SKILL.md — logging the format-convert event