mt-dialogues-100k-v1

Multi-turn medical dialogues — 100k_V1

39,996 synthetic doctor/patient dialogues, generated from EPFLiGHT/fully-open-meditron for supervised fine-tuning. Each one turns a single question-and-answer row of the source corpus into a conversation where a clinician has to elicit the case before answering it.

Files

The cases are included so the run can be reproduced, re-scored, or regenerated with another model. A dialogue file on its own cannot be.

How a dialogue is produced

Three participants, isolated by construction rather than by instruction:

• doctor — sees the opening line and the question it must answer. Nothing else. It chooses each turn whether to ask, to examine, or to conclude, and deciding it has enough is the capability being taught.
• patient — holds the spoken facts and discloses one per reply.
• records — holds the observed facts and releases them only on an examination or a chart lookup.

A third class of fact, withheld, goes to nobody: it is the question author's voice, and disclosing it would hand over the answer.

The doctor tags each turn so the harness can route it, and the tag is stripped before the turn is recorded. No benchmark syntax appears anywhere: the evaluation targets each impose their own protocol at test time, and training on one of them would be training on the benchmark.

Fields

messages is the conversation in chat format — the doctor is assistant, both the patient and the records desk are user. Alongside it travel the provenance (item_id, source_component, frame, question, gold_text, answer_summary) and every measurement taken during generation (turns_used, pivotal_disclosed_est, concluded_by_doctor, tag_retries, fabricated_results, ...), so a filter can work from this file alone.

Generation

What the run measured

The elicitation figure is a lower bound: it matches facts lexically against what the patient and the records desk actually said, so a fact disclosed in paraphrase is not counted.

Limitations

• Generated by a model, and not reviewed by a clinician. The final turn is the doctor's own conclusion, not the source corpus's expert answer — it can be wrong. gold_text and answer_summary are carried alongside so correctness can be scored, but no such filter has been applied here.
• Nothing is filtered. This is the raw generation output, on purpose: filters are scorers over the recorded measurements, and applying them here would destroy the ability to compare them.
• Inherits whatever biases and errors the source corpus holds.

Frames

Source components