HOBO docs
Changelog
Newest first.
2026-11-10: docs
- Removed a refusal figure measured on Do-Not-Answer. Its license is non-commercial, and these docs use only data we may use commercially. HOBO never trained on it.
2026-10-07: the current preview
- A new preview checkpoint, trained on more kinds of decision (policy cases, records, larger tool menus, when to call, ask or decline) and more ways of asking the same question. Every number in these docs is from it.
- It missed two tests we registered before training it, and we published it anyway: it reads90.0% on SQuAD, under the 92.2% we required, and on decision types it never trained on it averaged 48.5% against the earlier preview's 53.7%. It was never wrong while 90%+ sure on either. See Where HOBO struggles.
- Claims. FEVER 87.9%, against 81.4% for the earlier preview, with one claim calibration served on supports-or-refutes questions (70.8% without it).
- Fewer confident mistakes on real tool calls: 1.2% of answers wrong while 90%+ sure, against 2.3% for the earlier preview, at 87.7% right (88.8% before).
- Pressure and prestige. Told the user is sure of a wrong answer, it switches a right answer to it in0.7% of cases (1.8% before). The test the earlier preview missed, swapping the source on hard claims, it now passes.
- Korean needs no adjustment now: 83.6% of 3,000 KLUE-NLI claims, wrong while 90%+ sure on 1.3%.
2026-10-03: the second preview
- A preview checkpoint trained to hold its answer under pushback and to judge the evidence rather than its source.
- It missed one test we registered before training it, and we published it anyway: on hard claims, swapping the source changed 35 answers against 32 for a neutral rewording.
- Korean claim withdrawn, then earned back the same day: its first Korean read was wrong while 90%+ sure on 6.1% of a Korean claim test, over the 5% bar; a confidence adjustment fitted on separate practice data brought it to 0.2%.
- Pushback: switches under user pressure fell to 1.8%, from 7.9%. Quoted claims fixed: a claim put in someone's mouth reads as "supported" 0 of 300 times, from 83 to 93%.
2026-10-02: the first preview
- A checkpoint passed every bar registered before its training run.
- New: HOBO asks first when a request fits a tool but leaves out a detail it needs.
- HOBO is offered through an API, and as a private install under contract. No public download.
- An 8-bit build was checked for private installs. See Models.
- HOBO Mini archived. Its page stays up at /hobo-mini.
- Docs split into pages.
Next: more training
We'll keep training until HOBO's core behaviors are strong: confidence you can trust, handing back what it can't settle, and judging the evidence, not the source. Each run is registered before it starts and its results land here. See Future work.