Terminal
Terminal/Skill
Interview-based planning for Fireworks fine-tuning that picks a cookbook example, dataset and eval plan before any job is created.
By Fireworks AI
Works with
npx skills add fw-ai/cookbook --skill researchResearch: explore method, data, evaluation, and cookbook starting points.
Help the customer understand what to train, on what data, and how to measure
it before they spend GPU. This skill is the interview-driven front door (like
Tinker research). Execution lives in configure; failures live in debug.
| Research answers | Configure answers (later) |
|---|---|
| Implied method (SFT, DPO, RFT, embedding) from supervision signal | Exact method confirmation + hyperparameters |
| Cookbook entry — case study, example, or recipe path | Workflow surface (firectl vs SDK vs Training API) |
| Dataset plan — local, bundled, HF candidates, labeling schema | Upload, column mapping, job create |
| Eval plan — metric class, baseline, cookbook eval hook | Run baseline eval, wire evaluators |
| Suggested path (managed, serverless, dedicated) — coarse | Cost, model choice, monitor, deploy |
Research does not pick a final base model, estimate cost, or create jobs. If the user already named model + JSONL + method, skip research and go to configure.
Order: research → configure → (debug if needed).
Follow references/methodology.md. In short:
references/cookbook-catalog.md and
open the closest README before asking questions the catalog can answer.references/interview-questions.md;
propose options grounded in what you read; let the user revise.references/interview-questions.md passes.firectl.export FIREWORKS_SESSION_ID="$(python3 -c 'import uuid; print(uuid.uuid4())')"
export FIREWORKS_CLIENT_SOURCE="fireworks-training-skill/2.2.0"
Record entry_skill: research when writing a run manifest.
Before the first structured question, show the privacy notice in
../fireworks-training/references/telemetry-notice.md.
Check firectl skill-journey record --help once. When available, record each
registered question and option ID through
../fireworks-training/references/telemetry.md. Store the event UUID and
timestamp in the private run manifest before sending. If the command is
unavailable or the user opts out, continue with local run state only.
Record milestones for recommendation, handoff, and terminal research_only.
Never send raw customer prose, datasets, credentials, or paths.
If firectl whoami fails, tell the user to set a scoped API key with hidden
terminal input. Never ask for it in chat.
For a vague first message, show the privacy notice, then offer Research,
Configure, and Debug in the welcome-entry AskQuestion. Ask one question, then
stop. Record the registered welcome option after the answer.
Continue directly when the user clearly asks to explore a task, method, dataset, evaluation, or example. The privacy notice still precedes the first research interview question.
First line of every response (see references/output-template.md):
**Research** — exploring method, data, eval, and cookbook starting points.
Read references/cookbook-catalog.md. Open the best candidate README or example
README. Run references/interview-questions.md — one AskQuestion per turn.
STOP and wait. Do not recommend until the completion gate passes.
Use references/output-template.md:
Fire the Handoff AskQuestion. On plan_configure, write the handoff block
from references/cookbook-catalog.md into fireworks-training-runs/<run-id>/run.md,
then continue in configure.
Never tell the user to paste a canned configure prompt.
| Need | Reference |
|---|---|
| Full cookbook index | references/cookbook-catalog.md |
| Interview scripts and completion gate | references/interview-questions.md |
| Methodology (inspect → interview → eval) | references/methodology.md |
| Turn shapes and readiness package | references/output-template.md |
| Case study slugs (subset) | references/case-studies.md |
| Telemetry and privacy | ../fireworks-training/references/telemetry.md |
| Signal | Route to |
|---|---|
| User wants to train, deploy, or resume with known plan | configure |
| User reports failure, stuck job, or error | debug |
| User still exploring data, eval, or cookbook | stay in research |