format-adapter-sharegpt

Convert canonical training example records (@agentic/code/frameworks/training-complete/schemas/example-record.yaml) into ShareGPT-format JSONL — the multi-turn conversation schema popularized by Axolotl, LLaMA-Factory, and the open-source dialogue tuning community.

When to Use

• Emitting multi-turn dialogue task examples where turn roles matter
• Axolotl training runs (ShareGPT is Axolotl's preferred format for conversations)
• Preserving multi-turn structure that Alpaca flattens away

Parameters

• <input-glob> (required) — glob of canonical records
• --output <path> (optional) — default: .aiwg/training/exports/sharegpt-<timestamp>.jsonl
• --validate-round-trip (optional) — reload output and verify invariants

Format Spec

One JSON object per line with a conversations array of {from, value} turns:

json
{"conversations": [{"from": "system", "value": "You are helpful."}, {"from": "human", "value": "Hello"}, {"from": "gpt", "value": "Hi there!"}]}

Role mapping: system → "system", user → "human", assistant → "gpt".

Operation

1. Load canonical records — validate against schema.
2. Transform — build conversations array:

• input.system → {from: "system", value: ...} (if present)
• input.user → {from: "human", value: ...}
• output.assistant → {from: "gpt", value: ...}
• Multi-turn: traverse context_refs chain to reconstruct prior turns if any.

1. Validate target — at least one human + one gpt turn; role sequence alternates sensibly; Axolotl-compatibility check.
2. Round-trip check (if --validate-round-trip) — rebuild canonical record from conversations + sidecar.
3. Write output + log — emit JSONL, write sidecar, append format-convert event.

Round-Trip Invariants

Conversation structure captures input.system, input.user, output.assistant. Preserved via sidecar: id, task_type, full metadata, reasoning_trace. Tool calls are inlined into the gpt turn value (serialized JSON) when small; otherwise routed to sidecar.

Sidecar Metadata

<output>.metadata.yaml holds per-line: id, task_type, full metadata.*, output.reasoning_trace, large output.tool_calls, and any context_refs that could not be reconstructed as prior turns.

Acceptance Criteria

• Every canonical record emits a valid ShareGPT conversation (or is rejected with a logged reason).
• Multi-turn sequences preserve role ordering per Axolotl conventions.
• --validate-round-trip reconstructs all round-trip invariants.
• format-convert event logged with input/output/rejection counts.

References

• REF-472 — ChatML / ShareGPT / OpenAI messages format comparison
• ADR-022 D7 — canonical + adapter strategy

Delegation

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