Evaluating local models on real prompts, not demo trivia

Curate prompts from production-shaped traffic

Strip PII and secrets, bucket by intent (summarise, classify, translate), and include long-tail edge cases that burned you before. Avoid only testing limericks and prime-number riddles—they correlate poorly with assistant quality.

Scorecards beyond BLEU

Human rubrics help for small sets. For larger sets, use LLM-as-judge cautiously—bias and leniency are real. Prefer deterministic checks when outputs are structured JSON with parseable fields.

Latency and RAM together

A model that fits in VRAM on a desktop may not fit on a Pi. Measure tokens per second and peak RSS on target hardware. Thermal throttling changes week-two numbers.

Regression gates

When swapping model IDs, require a minimum pass rate on the eval set before promoting. Keep old prompts versioned with git so you can bisect quality drops.

Open weights and governance

Local models still need license review for redistribution if you ship images. Document provenance for compliance—even homelab projects sometimes graduate to team use.

Keep a changelog of quantisation formats and runtime versions; GGUF and friends evolve quickly, and silent speed regressions show up when you rebuild containers without pinning.

Share anonymised failure buckets with teammates: “hallucinated JSON keys” and “wrong language” deserve separate tracks so you do not tune prompts for the wrong failure mode.

Operational wrap-up: shipping without regret

When you operationalize the ideas behind “Evaluating local models on real prompts, not demo trivia,” start with a single toggle—an environment flag, a config stanza, or a feature branch deploy—that lets you compare old and new behaviour side by side. Use staging hardware you can afford to break: a spare Raspberry Pi, an old laptop, or a tiny cloud VM. Measure resident set size, cold-start time, p95 latency to your LLM provider, and error counts from journald or container logs before you point production webhooks at the stack. Week-one reviews usually surface missing timeouts, naive retry loops, and logging that omits request IDs; week-four reviews catch slow leaks, SD card exhaustion, and TLS renewal gaps. Write rollback steps next to rollout steps: which systemd unit to restore, which container tag to pin, which API key to rotate if a webhook secret leaks. Reliability is the product feature nobody applauds until it disappears.

Documentation debt kills homelab automation faster than clever bugs. For slug “evaluating-local-models-real-prompts,” keep a one-page runbook: ASCII diagram of data flow, listening ports, file paths for configs, and where secrets live on disk. Note the exact PicoClaw or companion binary version you deployed and link to upstream release notes. When vendors deprecate endpoints or models, you diff your runbook against official docs instead of archaeology on live systems. If anyone else—family, teammates—might restart services, document safe stop/start order and how to verify health. The goal is that a tired operator at midnight can follow steps without reading the entire blog archive.

Treat cost and reliability as one system: log every LLM call with approximate token counts, bucketed by workflow, and compare against invoices weekly. Spike detection should trigger investigation before budgets hard-fail—often a runaway cron or a duplicated webhook is the culprit, not “the model got smarter.” Pair financial telemetry with synthetic probes: a canary prompt that runs hourly and asserts latency and format constraints. When probes fail, page or notify through the same Telegram or Discord channels your humans already watch so anomalies do not live only in Grafana. This closing loop—money, latency, correctness—is how lightweight assistants remain boring infrastructure instead of science fair exhibits.

Where to go next in the PicoClaw knowledge base

This site’s guides translate patterns into commands: Raspberry Pi and Pi 5 setups, self-hosted assistants, Docker and Compose, systemd services, nginx HTTPS, Cloudflare Tunnel, Tailscale, n8n webhooks, Linux cron jobs, Telegram and Discord bots, and local models via Ollama or OpenAI-compatible gateways. The providers and configuration pages list how to wire OpenAI, Anthropic, Gemini, Groq, DeepSeek, OpenRouter, and more without scattering secrets across shells. Security, workspace, heartbeat, and API references explain sandboxing, scheduled prompts, and HTTP integration in depth—use them when you promote experiments to always-on services.

Comparison and alternatives articles situate lightweight Go agents next to heavier Python or Node stacks so you pick runtime deliberately, not by default. News and community links track upstream changes. If you are uncertain, ship the smallest vertical slice: one scheduled summary, one chat command, or one signed webhook—prove observability and cost discipline before layering complexity. Edge constraints on RAM, thermals, and bandwidth are not temporary hurdles; they define the niche where small binaries and clear policies outperform monolithic demos that never leave a developer laptop.

Finally, revisit this article—“Evaluating local models on real prompts, not demo trivia”—after your first production month. Annotate what aged poorly: a provider price change, a deprecated API field, a Pi firmware quirk. Update your internal notes and, if you maintain a public fork or gist, refresh it too. The niche moves quickly; static knowledge rots. PicoClaw’s model is to stay small at the edge while models and prices churn in the cloud—your documentation should echo that split: stable operational procedures on the left, volatile model cards on the right. Close the loop with metrics: dollars spent, incidents avoided, minutes saved. Those numbers justify the next iteration of your assistant better than any manifesto.

Accessibility and clarity matter even for personal bots: use descriptive command names, consistent help text, and error messages that suggest the next corrective action. Internationalization may not be your day-one priority, but encoding and emoji handling in chat bridges trips many newcomers—test with non-ASCII samples early. Backups of configuration and prompt templates belong in the same lifecycle as code: versioned, reviewed, restorable. These habits compound; they are how assistants remain maintainable when you are not the only operator anymore.

Performance tuning is iterative: profile before optimizing, and optimize the bottleneck you measured—not the framework you dislike. Network RTT to LLM endpoints often dominates; caching embeddings or deterministic template fragments locally can shave recurring costs. CPU spikes on Pis may be thermal or power-supply sag; rule those out before rewriting code. When you change models, re-benchmark end-to-end latency and weekly spend; a “smarter” model that doubles latency can break chat UX even if quality improves. Keep a changelog of model IDs and prompt hashes so regressions are bisectable instead of mysterious.