ugrad model lab

standalone desktop app real backend inference dataset + report export

Company IP: qbitOS / ugrad.ai · Author: Tad R. Ericson · Licensing: [email protected] · [email protected]

Compliance Hub Release Evidence Security.txt Compliance Repo Path ModelLab Source

Quantum Prefixes: pending

DAC: pending

Steno: pending

Preflight: pending

Iron Line tie-through: pending

Search/History/Freya links: pending

Iron Line refs Preflight refs History refs Search refs Freya refs

Model Lab Chat Prompt

asks current provider/model directly

Ask a question. I will run it with your active setup and show response + telemetry.

Question

Guided Cold Model Flow

step 1 of 6

1) Configure model + providerpending

Pick provider, set base URL, and model IDs. Use auto-calibration for your offline-first lane.

Provider

Auto-calibration profile

Base URL

Primary model id

Race model A

Race model B

Race model C

Race model D

Race uses same provider, base URL, prompts, and parameters for fair comparison.

Engine tag

API key (optional)

Temperature

Max tokens

2) Add simple question setpending

Use one question per line in Prompt suite (10-20 prompts works best for races).

Prompt pack

Prompt suite (one prompt per line)

3) Run suite (cold pass)pending

Run one cold pass, then race A/B/C/D for direct comparison.

Minor tweak: prompt prefix

Minor tweak: prompt suffix

Minor tweak: temperature delta

Minor tweak: max tokens delta

Iteration note

iteration 0

model lab ready

4) Export benchmark reportpending

Export report JSON to compare model runs (latency, score, throughput).

5) Export Q/A datasetpending

Export JSONL for fine-tune/eval corpus and replay tests.

6) Iterate + save best pairpending

Save best pair, open chat for interactive tests, or reset and rerun.

dataset rows: 0 last report: none provider health: unknown