moganai/lavoir-dialogues

Lavoir Dialogues

Customer-service conversations for learning when to ask a clarifying question and which one.

32,412 synthetic conversations from 12 English customer-service workflows (banking, e-commerce returns, HR, insurance claims, IT help desk, privacy requests, telecom, travel changes, and 4 held-out workflows). Each example is a routing decision (which team should handle this?) together with everything needed to learn and evaluate a question-asking policy:

• the exact posterior over teams given what the customer has said so far (the soft training target),
• the true expected value of information (VOI) of every question that could still be asked,
• a probe for every question: the question text and a realistic customer answer, so any policy can be replayed without a live user.

The routing rules are written by hand, so every label and every VOI is exact rather than annotated. An LLM only turns hidden facts into natural customer messages and answers, and a second LLM from another family checks that no hidden fact leaked into the text. This is the question-asking training and test data of the Lavoir model.

Workflows and splits

workflow	split	units	decisive / all slots	train	test_seen	test_zeroshot
banking_support	seen	5	4 / 5	2,546	619	–
ecommerce_returns	seen	5	4 / 5	2,787	698	–
hr_requests	seen	6	3 / 4	2,762	667	–
insurance_claims	seen	5	4 / 5	2,768	680	–
it_helpdesk	seen	5	4 / 5	2,678	682	–
privacy_requests	seen	5	4 / 5	2,593	680	–
telecom_support	seen	6	4 / 5	2,732	678	–
travel_changes	seen	5	4 / 5	2,735	675	–
clinic_routing	unseen	6	4 / 5	–	–	1,370
parcel_delivery	unseen	5	4 / 5	–	–	1,351
saas_support	unseen	5	4 / 5	–	–	1,373
student_affairs	unseen	6	4 / 5	–	–	1,338
total				21,601	5,379	5,432

• train and test_seen use the same 8 workflows with different conversations (different random seeds).
• test_zeroshot has 4 workflows that are not in train at all.
• Every workflow is a separate config because each one has its own teams and slots: load_dataset("moganai/lavoir-dialogues", "banking_support").
• With datasets, dict fields share one schema per workflow, so slots that were already asked show up as None in slots, oracle_voi and probes (the raw JSONL simply leaves them out). Drop None values before use.

What an example looks like

{
  "id": "train/ecommerce_returns/00001/p2/b",
  "workflow": "ecommerce_returns", "split": "train", "kind": "b",
  "state": [
    {"role": "user", "text": "Hi. My name is Daniel. I write to you because I need some help with something. I would like to have money back. ..."},
    {"role": "system", "text": "When was it delivered: within the last 30 days or earlier?"},
    {"role": "user", "text": "I received it more than 30 days ago."}
  ],
  "question": {"type": "choice", "instructions": "Which team should handle this order problem?",
               "criteria": {"returns_desk": "returns desk: standard returns and exchanges",
                            "refunds": "refunds team: money back for faulty or wrong deliveries", "...": "..."}},
  "target": {"returns_desk": 0.21, "refunds": 0.315, "logistics": 0.0, "marketplace_support": 0.30, "warranty_service": 0.175},
  "gold": "marketplace_support",
  "profile": {"problem": "stopped_working", "seller": "third_party", "delivery_age": "over_30_days", "wants": "money_back", "customer_name": "Daniel"},
  "slots": {"problem": "what is wrong with the order", "seller": "who sold the item", "wants": "what the customer would like", "customer_name": "the customer's first name"},
  "asked": ["delivery_age"], "answer_kind": "clean",
  "evidence": {"wants": ["money_back"], "delivery_age": ["over_30_days"]},
  "oracle_voi": {"problem": 0.035, "seller": 0.5425, "wants": 0.035, "customer_name": 0.0},
  "probes": {"seller": {"q": "Was the item sold by us directly or by a seller on our marketplace?",
                        "a": "The item was sold by a third party.", "kind": "clean", "reveal": {"seller": ["third_party"]}}, "...": "..."},
  "style": "written by a non-native English speaker, simple words"
}

After one answer, four teams are still possible and asking who sold the item is worth +0.54 in the probability of the correct team; asking the customer's name is worth nothing.

field	meaning
state	the conversation so far: user turns are the customer, system turns are the questions asked
question	the routing decision in Laya's typed format (criteria: team → description)
target	exact posterior over teams given evidence: the soft training target
gold	the team the hidden profile actually belongs to
profile	the hidden ground truth: one value per slot
slots	questions that can still be asked (name → description); already-asked slots are removed
asked, answer_kind	slots asked so far and the kind of the last answer: clean, partial ("it might be A or B"), unknown ("I don't know") or extra (also reveals an unasked slot)
evidence	what the conversation has revealed: slot → set of values still possible
oracle_voi	exact expected increase in p(gold) if that slot's question is asked now
probes	per slot: question q, the customer's answer a sampled from the answer model, its kind, and what it reveals
kind	a = first message only, b = after one question, c = after two
style	writing style asked of the generator (10 styles: short, angry, typos, bullet points, formal e-mail ...)

The schemas/ folder has the 12 workflow definitions: teams, slots with their values, priors (with at most one parent slot), the ordered routing rules, and the question text of each slot. target and oracle_voi can be recomputed from them exactly.

How it was made

1. Schemas. Each workflow has 5–6 teams, 3–4 decisive slots and one non-decisive slot (the customer's name). Automatic gates checked that every team is reachable, team masses lie in [0.03, 0.5], at least 5% of the mass needs two or more slots, and a dry run of the exact posterior is neither too easy nor too hard.
2. Cases. A case is a partial profile (what the first message states) with 2–4 complete profiles consistent with it, which usually differ in their correct team. Each profile gives a first-message example (a), with probability 0.7 one after a question and answer (b), and with probability 0.4 one after two (c).
3. Answers are clean (67.5%), partial (12.5%), "I don't know" (10%) or over-informative (10%).
4. Text. Qwen3.6-35B-A3B writes the messages (10 styles, at most 120 words) and answers; its prompt lists the facts to state and the topics not to mention, never their values. Gemma-4-26B-A4B then guesses every decisive slot from the text with an evidence level. A text is regenerated (up to 3 times) if a hidden slot can be read from it (leakage) or a stated slot cannot (fidelity); cases that fail three times are dropped.
5. First-attempt message rejection was 1–10% for 7 training workflows and 22% for privacy requests, whose "type of data" and "request" slots overlap; answer rejection was at most 2.8%; partial answers were guessed correctly 46–62% of the time (50% expected).

Uses

• Train a model that returns a decision and the value of each candidate question, e.g. with the Lavoir code: python -m lavoir.train --phase decision --data "data/train/*.jsonl" ... (the files are in the format the training code reads).
• Evaluate question-asking policies against an exact oracle: replay each a example, answer every question with its probe, and compare with oracle_voi.
• Calibration studies: target is the true posterior, not a one-hot label.

Limitations

• Synthetic: rules, priors and answer kinds were chosen by us, and LLM-written customers are more cooperative and more on-topic than real ones.
• Dropping cases whose text failed the checks three times is a mild selection effect: 0–3% of the cases per workflow, 5% in privacy requests.
• English only; the Turkish counterpart is lavoir-dialogues-tr.
• Names are fictitious and sampled from a fixed list; no real customer data was used.

License

CC BY 4.0. The texts were generated with Qwen3.6-35B-A3B and checked with Gemma-4-26B-A4B, both Apache-2.0.

Citation

@misc{yilmaz2026lavoir,
  title        = {Lavoir: A Single-Pass Decision Model That Knows Which Question to Ask},
  author       = {Furkan Yilmaz and Habibe Aleyna Tasdemir and Muhammed Faruk Gozay},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/moganai/lavoir}}
}

---

Support MoganAI
If our open models and datasets are useful to you, you can support our work.
Açık modellerimiz ve veri setlerimiz işinize yarıyorsa çalışmalarımıza destek olabilirsiniz.