Serverless Training quickstart: train a support agent with Serverless RL

[dbrian57]

What this adds

The Serverless Training section previously had no in-house getting-started content — the "How to use" pages linked out to OpenPipe ART's external quickstart, whose example task (the 2048 game) is academic. This PR adds a self-contained quickstart built around one of the product's officially supported use cases: customer support.

• serverless-training/quickstart.mdx — trains a LoRA for a customer support agent (fictional smart thermostat) with Serverless RL: register OpenPipe/Qwen3-14B-Instruct with ART's ServerlessBackend → tool-calling rollout over an inline 10-article knowledge base → RULER scoring (no labeled data or hand-written reward function) → 6-step training loop → inference against the trained LoRA via the wandb-artifact:/// endpoint.

[w-b-hivemind]

HiveMind Sessions

1 session · 43m · $19

Session	Agent	Duration	Tokens	Cost	Lines
Serverless Training LoRA Quickstart Documentation 3223f79f-743d-4d6c-8403-04d76dfe5c06	claude	43m	87.4K	$19	+405 -6
Total		43m	87.4K	$19	+405 -6

View all sessions in HiveMind →

Run claude --resume 3223f79f-743d-4d6c-8403-04d76dfe5c06 to pickup where you left off.

[dbrian57]

Sources and decision log

Sources for factual claims

• Product behavior, endpoints, pricing, W&B services used: existing pages in this repo — serverless-training.mdx, serverless-training/usage-limits.mdx, serverless-training/use-trained-models.mdx, serverless-training/available-models.mdx, serverless-training/api-reference.mdx (base URL https://api.training.wandb.ai/v1, wandb-artifact:///ENTITY/PROJECT/MODEL:STEP endpoint schema, free training during preview).
• ART API usage (TrainableModel, ServerlessBackend, model.register, model.inference_base_url / inference_api_key / get_inference_name, art.Trajectory, gather_trajectory_groups, iterate_dataset, backend.train, model.log, ruler_score_group): OpenPipe's current ART·E notebook (openpipe/art-notebooks/examples/art-e.ipynb, pins openpipe-art==0.5.9) and art.openpipe.ai docs (art-client, installation-setup, ruler, sft-training pages).
• Use-case selection (customer support as an officially supported use case): internal launch messaging (Slack #revenue launch thread 2025-10-08; PMM guidance names voice agents, customer support, agentic RAG, deep research). Internal docs-plan context: Slack #serverless-training thread 2025-10-01 (launch docs intentionally barebones; eng lead open to new content); JIRA DOCS-2698 (Serverless Training umbrella naming/IA).
• Base model choice: serverless-training/available-models.mdx catalog (auto-generated).

Key decisions

• Customer support agent instead of adapting ART·E or 2048: officially supported use case; self-contained inline knowledge base (no Enron dataset download); demonstrates the "no labeled data" RULER story. Confirmed with Dan Brian before drafting.
• Full train→serve loop on one page: internal launch threads agreed inference-with-LoRAs must be documented with training; the page ends at the wandb-artifact inference call and links to Use your trained models.
• Manual tool JSON schema instead of langchain_core.convert_to_openai_tool: drops three dependencies from the reader's environment vs. the ART·E notebook.
• No validation split / checkpoint deletion in the loop: kept the quickstart minimal; both are pointed to in Next steps (ART·E notebook, checkpoint-deletion docs).
• No wall-clock/cost estimate: intentionally omitted (confirmed with Dan) — no measured number exists for this exact configuration.

Intentionally omitted

• Serverless SFT walkthrough (mentioned only in Next steps; candidate for a follow-up quickstart).
• Weave integration in the main code path (optional Tip instead, to keep required deps at openpipe-art only).

Needs SME verification

• End-to-end execution: nobody has run the full training loop yet. Dan Brian will run the Colab; verify rollouts, RULER calls, and training steps complete against the current backend.
• Model catalog freshness: page uses OpenPipe/Qwen3-14B-Instruct from our published catalog, but OpenPipe's current ART·E notebook trains Qwen/Qwen3.6-27B, which isn't in available-models.mdx — is the auto-generated catalog current?
• Checkpoint alias: final example serves ...support-agent-001:step6. Confirm a 6-step run labels its final checkpoint exactly step6 (format inferred from use-trained-models.mdx).
• RULER judge: quickstart requires an OpenAI key for the judge (openai/gpt-5.4, matching OpenPipe's notebook). Is a W&B Inference-hosted judge supported through ruler_score_group? If so, a single-key story would onboard better (follow-up).
• iterate_dataset/backend.train signatures: match openpipe-art 0.5.9; confirm the pinned version customers get from a bare pip install openpipe-art hasn't drifted.

Suggested reviewers: David Corbitt (technical accuracy), John Mulhausen (docs IA/editorial).

Resume prompt

Continue work on wandb/docs PR #2879 (branch dbrian/docs-serverless-training-quickstart): the Serverless Training quickstart (serverless-training/quickstart.mdx + serverless-training/cookbooks/train-support-agent.ipynb). Decisions already made: customer support agent use case, full RL loop with RULER (OpenAI judge, openai/gpt-5.4), base model OpenPipe/Qwen3-14B-Instruct, page + Colab companion, no time/cost estimate. Remaining: (1) apply the style-guide pass files (bundled in the docs plugin skill directory) and push fixes; (2) incorporate results of Dan Brian's end-to-end Colab run, especially the actual checkpoint alias for step 6; (3) resolve the SME checklist above; (4) mark ready for review and assign David Corbitt + a TW reviewer. Note: the local repo's css-tab-borders branch has unrelated uncommitted CSS changes — use a git worktree for commits instead of switching branches.

{
  "skill": "docs",
  "version": "unknown",
  "depends_on": "unknown",
  "note": "versions.json not present in the installed skill directory (cw-claude-code-plugins/docs/1.0.0/skills/docs); versions not guessed.",
  "model": "claude-fable-5"
}

[dbrian57]

Style-guide pass complete (commit 5bfa166): descriptive headings without step numbering, bracket-style placeholders, complete-sentence list intros. Remaining before marking ready: end-to-end Colab run (Dan) and the SME checklist above.

[github-actions]

📚 Mintlify Preview Links

🔗 View Full Preview

✨ Added (1 total)

📄 Pages (1)

File	Preview
serverless-training/quickstart.mdx	Quickstart

📝 Changed (4 total)

📄 Pages (3)

File	Preview
serverless-training.mdx	Serverless Training
serverless-training/prerequisites.mdx	Prerequisites
serverless-training/usage.mdx	Usage

⚙️ Other (1)

File
docs.json

---

🤖 Generated automatically when Mintlify deployment succeeds
📍 Deployment: 5bfa166 at 2026-07-08 21:41:22 UTC

[github-actions]

🔗 Link Checker Results

✅ All links are valid!

No broken links were detected.

Checked against: https://wb-21fd5541-dbrian-docs-serverless-training-quickstart.mintlify.app

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
wandb	🟢 Ready	View Preview	Jul 8, 2026, 9:41 PM