Skip to content
posttrainllm docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

`posttrainllm quickstart` — your first specialist in one command

posttrainllm quickstart — your first specialist in one command

quickstart turns a data file into a trained, runnable specialist on your Mac with one command and zero ML knowledge: it inspects the data, auto-picks a base from the gallery, infers a LoRA recipe, trains, and samples the result so you can see whether it helped. It is the CLI sibling of the Mac app’s Factory tab (B6) and shares its decision core (RecipeResolver) with it.

Everything runs on-device. No account, no cloud upload — the only network is the initial base-model pull.

One command

posttrainllm quickstart mydata.jsonl --yes

That inspects mydata.jsonl, resolves a (base, recipe), trains an adapter to adapter.lora, writes a reproducible posttrainllm.project.json, and prints a few sample completions from the new specialist.

See the plan without training first:

posttrainllm quickstart mydata.jsonl --dry-run

What it accepts

quickstart detects the data shape from the first lines:

Your data looks like Detected shape What it does
{"messages":[{role,content},…]} chat LoRA fine-tune, chatml template
messages with tool_calls / a "tools" key toolCall LoRA fine-tune, longer --max-seq
{"instruction","output"} or {"prompt","completion"} instruction LoRA fine-tune
not JSON (a plain-text corpus) rawText from-scratch pretrain (use posttrainllm train)

If it can’t classify the data it tells you the expected formats and exits non-zero rather than guessing silently.

What it picks

  • Base — the smallest gallery model whose tags match the data shape (e.g. a tool/agent-tagged base for tool-call data), preferring smaller models so it fits a laptop. Override with --base <gallery-id | path | hf-id>.
  • Recipe — LoRA rank and step budget scale with dataset size (<500 rows → r8/300 steps, <5000 → r16/800, else r32/1500); α = 2·rank, lr 2e-4, sequence packing on, NEFTune on. The resolved recipe is printed before any training so you can eyeball it.

The full mapping lives in RecipeResolver (see native-mac/Sources/TinyGPTModel/RecipeResolver.swift) and is unit-tested.

Output

  • adapter.lora — the trained adapter (--out to change).
  • posttrainllm.project.json — base + adapter pins so the result is reproducible and shippable (the B31 project-pin format; passes posttrainllm validate).

Flags

Flag Meaning
--dry-run print the resolved plan + project file; train nothing
--base <id|path|hf-id> override the auto-picked base
--gallery <path> gallery manifest.json (default: ./gallery/manifest.json)
--out <path> adapter output (default adapter.lora)
--samples <N> demo samples after training (default 3)
--yes, -y skip the train confirmation

Verify (no GPU)

bash evals/quickstart-smoke.sh

Asserts the --dry-run plan contract against fixture data: chat data picks the chat base, tool-call data picks the tool base with max-seq=2048, the project preview carries an adapter pin, and a missing file exits non-zero.

Limits (V1)

  • From-scratch (raw-text) training isn’t wired into quickstart yet — it prints the plan but routes you to posttrainllm train.
  • Auto-pull of a bare gallery id isn’t wired: if the auto-picked base is a gallery id with no local weights, pass --base <local-path-or-hf-id> (or posttrainllm pull it first). Paths and HF ids train directly.
  • Single SFT pass — no SFT→DPO→quantize chains yet (the recipe resolver can grow stages later). See docs/prds/B33-laptop-finetune-onboarding.md.

Was this page helpful?