Med-Augment-GLM5

Synthetic supplemental training data for four medical tasks, generated with z-ai/glm-5.2 and independently reviewed both at generation time and via a second-pass adversarial audit.

Tasks

• med_dialog_healthcaremagic / med_dialog_icliniq — one-sentence faithful summaries of the patient's core question from real HealthCareMagic/iCliniq patient-doctor dialogues.
• pubmed_qa — yes/no/maybe answers to real PubMedQA research questions.
• medi_qa — grounded, consumer-facing 3-6 sentence answers to real consumer health questions, sourced against MedlinePlus-style reference material.
• race_based_med — synthetic clinical Q&A pairs labeled for whether the response relies on harmful/inaccurate race-based medical reasoning (eGFR, spirometry, VBAC, pulse oximetry, STS score, pain-perception bias, and related topics).

Splits

• train — rows that passed generation-time review, an independent post-hoc quality audit, and an automated truncation check (free-text outputs must end on a complete sentence).
• errors — rows dropped for a failed review verdict, a generation error, or a truncated output. Kept for transparency rather than discarded.

Per-task row counts (train)

Task	Rows
med_dialog_healthcaremagic	736
med_dialog_icliniq	324
pubmed_qa	857
medi_qa	211
race_based_med	317

Quality audit

A 50-row stratified sample (10/task) was independently re-judged by a separate model than the one that generated and reviewed the data. Verdicts: 34 GOOD, 13 MINOR_ISSUE, 3 MAJOR_ISSUE. All MAJOR issues were truncated outputs, which were then mechanically detected and moved to errors across the full dataset (49 of 1,619 free-text rows, ~3%). Remaining MINOR issues were mostly scope deviations (summarizing the full dialogue instead of just the core question) rather than factual fabrication, and did not warrant exclusion.

Fields

uid, task, orig_split, input, raw_input, context, ground_truth, thinking, output, model, finish_reason, self_answer_matches_gold, review_pass, review_verdict, error.