Bloom Tech Support — Field Resolutions Dataset

Quality, measured — not claimed

Here's what the score means in practice: take a real tech-support problem this company solved in the past. Hide that job's answer, then ask an AI — armed only with the rest of this dataset — to figure out the fix. About 1 in 3 times overall, and over half the time on recurring job types, the AI lands on the exact fix that actually worked — judged by an independent AI evaluator, not by us. That's the difference between a robot guessing from general knowledge and a robot that has effectively seen your problem before.

This is the Last Mile Exam, our quality evaluation for on-the-job datasets. The scored set stays private and refreshes with real jobs, so the exam can't be studied for. Full methodology available to licensees under NDA (publication pending our patent filing).

What a case file looks like

problem_category: Printer · resolved: fully_resolved · paid: true

PROBLEM: HP OfficeJet reports "paper jam" repeatedly; customer
finds no paper stuck anywhere. Restarting does not clear it.

RESOLUTION (worker-confirmed fix): Tray paper-size setting was
8x14 Glossy but Letter plain paper was loaded. Set paper type to
8.5x11 Letter, Plain — error cleared without any teardown.

Each file is a real resolved conversation — the fix is in the transcript, worked out live with the customer. On a growing subset, the technician's confirmed one-line fix is extracted to the top (shown above). Either way it's the part no scraped data has: what actually worked, in the field. PII is stripped before storage — de-identified by construction, with signed chain-of-title from the data owner.

Coverage

Computer repair · Phones & tablets · Wi-Fi/networking · Printers · Smart home · TV/streaming · Data recovery · Virus/malware · Setup & installation — the real distribution of in-home tech support work, not a synthetic benchmark.

Per-seat: a 1-tech shop pays 1 seat. Stack additional packs for 10–20% volume discounts. Included in your 14-day free trial. Cancel anytime. 60% of net fees go to the dataset owner — the model your corpus joins when you list.