Multi-Label Text Classification: One Choice, Several Yes/No Questions, or a Second Pass

Read the action, not the model brochure. “I was charged twice and the export now takes two minutes” is both a billing issue and a performance issue. If the next step is “hand this to the one team that should answer”, you want a single primary label: see intent-classification and text-classification-api. If the next step is “tag every applicable theme so a weekly report does not drop one of them”, you want multi-label. Forcing a Choice hides one of the two facts. This page owns that contract. It does not replace the single-request how-to, and it does not restate Make’s “assign only one category” switch — that control lives in Make, described in make-categorize-text. Customer feedback that only needs the main theme plus a priority stays on customer-feedback-classification.

Pick the shape your warehouse can store. The table is a contract map, not a ranking. No option removes the need to count false tags on your own labeled rows.

Write each Noul as one proposition: “Is a billing issue present?” Give criteria.true and criteria.false as short descriptions, 2 to 350 characters each, and instructions of 8 to 600 characters. Question ids follow the same rule as Choice: a leading letter, then lowercase letters, digits, and underscores, at most 40 characters. Do not ask two questions that mean the same fact, such as billing and invoice, or a true count will double. Skip a catch-all question unless someone will read it; a low probability is already a reason to abstain. The field noul is the probability of yes, from 0 to 1. There is no separate confidence field. A value near 0.5 is uncertain: send that tag to a person. Do not treat a high value on eight questions as eight confirmed tags until you measure false positives per tag. One request may mix one Choice (the primary route) and several Noul tags, as long as the total is 1 to 8 questions. Eight Noul questions are still one run. The example below is a request shape, not a recorded result. Thresholds come from your labels, not from the sample numbers you might imagine. See question-types and confidence-and-abstention.

curl --fail-with-body https://jevmodel.app/api/v1/systemone \
  -H "Authorization: Bearer $JEVMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"state":"I was charged twice and exporting a report now takes two minutes.","questions":{"billing":{"type":"noul","instructions":"Is a billing or charge issue present in this text?","criteria":{"true":"The text mentions charges, invoices, refunds, or payment.","false":"The text does not mention a billing issue."}},"performance":{"type":"noul","instructions":"Is a performance or speed problem present in this text?","criteria":{"true":"The text describes slowness, timeouts, or a speed regression.","false":"The text does not describe a performance problem."}}}}'

Do not flatten thirty labels into one Choice. The maximum is twelve, and a flat list that long makes the descriptions overlap. First request: a parent Choice that includes an explicit review option. Second request: only the child criteria for the parent that was chosen, and only when that parent’s confidence clears a threshold you measured. Otherwise a person picks the child. Each call is billed on its own. There is no idempotency key, so write the request into your history before you retry a lost response. If you need more than eight Noul tags, split them into two requests with the same state and disjoint question ids, and store both answers under a row id you generate. The two calls do not see each other. A Decision Flow can draw that branch in the console, but running a flow is for paid plans only: subscribe or buy a token pack. Flows do not use the daily free runs. See decision-flows. Every validated request, including the questions, must fit in 8,000 characters.

Do not collapse the result into one accuracy number. A comment with two true tags counts twice. For each tag, count false positives and false negatives on a held-out set, or report precision and recall per tag if you already compute them. A model that never predicts a rare tag can look strong on a single average and still drop that tag every week. Console Batch exports row-level answers and does not compute these counts; join them in your own script. Score a JSON list and a Hugging Face multi_label run on the same rows if those are the alternatives. Hugging Face scores are NLI entailment signals; reject any generated name that is not in your taxonomy before the join. The zero-shot path is zero-shot-text-classification. The held-out split and the habit of re-checking after a criteria change are evaluation. Free verified accounts share 5 successful runs per UTC day across Playground, each Batch row, and the API. A second pass is a second run.

It will not return an open list of tags you did not name. It will not read images, audio, or attachments. It is not a trained multi-label head you can download, and the product is not open source. Choice will not emit two winning labels from one question. Noul probabilities on different questions are not parts of one distribution, so they do not need to sum to 1, and they also must not be averaged into a single confidence. A generative JSON repair step is your code, not a field on this response. If the text is a raw email thread, clean it with the steps in email-classification before you ask several tag questions. If you only needed the primary product theme, go back to customer-feedback-classification rather than tagging every synonym.