Where you raise a parrot.
Bootstrap SFT corpus generation for Swift Parrot
(spoomplesmaxx-flash-35B-A3). Three data lanes feed one gated funnel and one
serializer:
- Lane A — agentic: teacher models run task-bank errands in-character as Olivia through hermes-agent batch datagen (pinned checkout, called, never forked).
- Lane B — fiction extraction: CoSER-style decomposition of organic fiction (your ebook library, AO3, Gutenberg) into character profiles, scenes, and dialogue with inner thoughts — the non-LLM entropy anchor and the theory-of-mind data.
- Lane C — self-play: user-simulator ⇄ character conversations tuned for SillyTavern turn dynamics (lazy typo'd user turns, length control, impersonation avoidance).
Everything is outcome-verified, rubric-judged cross-vendor, scrubbed, deduped,
span-safely harmonized, and rendered through a single serializer into Swift
Parrot's ChatML + Hermes <tool_call> template — with-thoughts and no-thoughts
variants, next-speaker-prediction samples, and a DPO artifact from judge-rejected
rollout siblings. The task bank and verifiers become the GRPO reward suite later.
- The contracts (start here, especially if you are Claude Code):
CLAUDE.md - The serializer byte format:
render/CONTRACT.md
uv sync
just test # offline suite: goldens, gates, lanes, guards
cp .env.example .env # keys: DEEPSEEK_API_KEY, GLM_API_KEY, OPENROUTER_API_KEY
# 1. verify teacher snapshot ids + prices (datagen/configs/teachers.yaml, pricing.yaml)
# 2. list books in datagen/configs/lane_b.yaml (local paths, never committed)
just pilot # small calibration run -> runs/<date>-pilot.manifest.yaml
just gate <run_id>
just render <run_id>
just stats <run_id>
just burn # refuses without a recent healthy pilotLane A additionally needs a pinned hermes-agent checkout: set
AVIARY_HERMES_DIR, run just install, add a to lanes: in
datagen/configs/pilot.yaml, and fill the task bank (tasks/).
Implemented and offline-tested end-to-end (fake-teacher fixtures; no live API calls in tests). Before the first paid run: verify dated teacher snapshot IDs and pricing, add books, and run a spend-capped lane B smoke on one book.
Training code (spoomplesmaxx repo), hermes forks/patches, trajectory data in
git (private HF dataset repos, referenced from runs/ manifests).