Pace-tuned planner v1 — training plan

Status: export wired — opt-in toggle in Settings → Models writes anonymized JSONL locally; copy into repo with scripts/export-pace-tuned-turns.sh. LoRA run still waits on enough collected turns.

Target

Base: mlx-community/Qwen3-4B-Instruct-2507-bf16
Output: pace-ai/pace-planner-v1 (HuggingFace + Sparkle manifest)
Ship path: RemoteModelManifestURL or Info.plist BundledMLXPlannerModelIdentifier bump

Dataset

Enable Settings → Models → Contribute anonymized planner turns (default OFF).
Use Pace — planner turns append to ~/Library/Application Support/Pace/pace-tuned-turns.jsonl (emails, phone numbers, home paths redacted). Now collects cloud/bridge (Codex, Claude) turns too, each tagged with meta.plannerProvenance — the distill-a-strong-teacher-into-our-own-model strategy. ⚠️ ToS-sensitive: OpenAI/Anthropic terms generally prohibit training a competing model on their outputs, so filter by plannerProvenance before training/shipping any model that would be distributed. Default ON (isPaceTunedTurnExportEnabled).
Copy into the repo: bash scripts/export-pace-tuned-turns.sh → evals/pace-tuned-export/export-YYYYMMDD.jsonl.
Mix with existing evals/fm-fixtures/*.txt converted to v10 JSON envelope shape.
Hold out evals/fm-fixtures-holdout/ — never train on holdout.

Train

bash scripts/train-pace-tuned-model.sh --check
# follow printed mlx_lm.lora command after dataset exists

Eval gate (must pass before default switch)

bash scripts/eval-v10-gate.sh
PACE_RUN_MLX_EVAL=1 bash scripts/eval-v10-gate.sh
python3 scripts/eval-planners.py --models <candidate-id>

Update PaceBundledModelsSettingsTests.shippingDefaults pin when the candidate wins.

Ship

bash scripts/train-pace-tuned-model.sh --emit-manifest pace-ai/pace-planner-v1 > remote-model-manifest.json

Host manifest → set RemoteModelManifestURL in Info.plist → Sparkle release.