med6m: a LoRA adapter for Qwen3-8B on medical multiple-choice QA — sealed-test evidence for checkpoint 443 at scale 0.5 / med6m: un adattatore LoRA per Qwen3-8B per QA medico a scelta multipla — evidenza su test sigillato

English. A rank-64 LoRA adapter for Qwen3-8B (unsloth bnb-4bit), QLoRA-trained on a commercially licensed medical mix (MedMCQA-train, MedQA-train, Italian and English medical Wikipedia) plus 22 % general replay. On a sealed, single-use test of 3,766 medical multiple-choice items (MedMCQA, MedQA, MMLU-medical, PubMedQA), at LoRA scale 0.5, accuracy rises from 61.34 % to 63.67 % (Δ = +2.34 pp, 95 % paired-bootstrap CI [+1.14, +3.53], exact McNemar p = 1.4·10-4; 306 items only the adapter answers correctly vs 218 only the base). MedQA (+1.55) and MMLU-medical (+0.74) are positive but not individually significant; the Italian ARC guard does not regress (+0.74 pp). Stated limits: the released artefact is checkpoint 443 of the second stage (not the final step 886); scoring is answer-letter log-probability only (no free-text generation, no safety or clinical validation); part of the MedMCQA gain is in-distribution; GSM8K and IFEval have not yet been measured on this checkpoint; the network audit is static only. Not a medical device; do not use for medical advice. The paper contains the formulas, figures, per-benchmark tables with CIs and tests, pinned hashes and the network-audit note. Italiano. Adattatore LoRA di rango 64 per Qwen3-8B, addestrato con QLoRA su un mix medico a licenza commerciale (MedMCQA-train, MedQA-train, Wikipedia medica IT/EN) più 22 % di replay generale. Su un test sigillato a uso singolo di 3.766 item medici a scelta multipla, a scala LoRA 0,5, l’accuratezza passa dal 61,34 % al 63,67 % (Δ = +2,34 pp, IC 95 % [+1,14; +3,53], McNemar esatto p = 1,4·10-4). MedQA e MMLU-med non sono singolarmente significativi; la guardia ARC italiana non peggiora. Limiti dichiarati: checkpoint 443 (non l’886 finale); solo log-probabilità della lettera; GSM8K e IFEval non ancora misurati; audit di rete solo statico. Non è un dispositivo medico: non usare per consigli medici. Files: article_en.pdf, article_it.pdf, code.zip (statistics, figures, LaTeX sources, original training/evaluation scripts, checksums). Adapter weights: Hugging Face odelpa/qwen3-8b-med6m; code: GitHub usergdp136/qwen3-8b-med6m. Licences: article, figures and adapter CC BY-SA 4.0; code Apache-2.0; base model Apache-2.0.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-02
DOI
https://doi.org/10.5281/zenodo.23895099
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

med6m: a LoRA adapter for Qwen3-8B on medical multiple-choice QA — sealed-test evidence for checkpoint 443 at scale 0.5 / med6m: un adattatore LoRA per Qwen3-8B per QA medico a scelta multipla — evidenza su test sigillato

Giovanni Del Papa
2 links
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
preprint

med6m: a LoRA adapter for Qwen3-8B on medical multiple-choice QA — sealed-test evidence for checkpoint 443 at scale 0.5 / med6m: un adattatore LoRA per Qwen3-8B per QA medico a scelta multipla — evidenza su test sigillato

Giovanni Del Papa
preprint en

Abstract

English. A rank-64 LoRA adapter for Qwen3-8B (unsloth bnb-4bit), QLoRA-trained on a commercially licensed medical mix (MedMCQA-train, MedQA-train, Italian and English medical Wikipedia) plus 22 % general replay. On a sealed, single-use test of 3,766 medical multiple-choice items (MedMCQA, MedQA, MMLU-medical, PubMedQA), at LoRA scale 0.5, accuracy rises from 61.34 % to 63.67 % (Δ = +2.34 pp, 95 % paired-bootstrap CI [+1.14, +3.53], exact McNemar p = 1.4·10-4; 306 items only the adapter answers correctly vs 218 only the base). MedQA (+1.55) and MMLU-medical (+0.74) are positive but not individually significant; the Italian ARC guard does not regress (+0.74 pp). Stated limits: the released artefact is checkpoint 443 of the second stage (not the final step 886); scoring is answer-letter log-probability only (no free-text generation, no safety or clinical validation); part of the MedMCQA gain is in-distribution; GSM8K and IFEval have not yet been measured on this checkpoint; the network audit is static only. Not a medical device; do not use for medical advice. The paper contains the formulas, figures, per-benchmark tables with CIs and tests, pinned hashes and the network-audit note. Italiano. Adattatore LoRA di rango 64 per Qwen3-8B, addestrato con QLoRA su un mix medico a licenza commerciale (MedMCQA-train, MedQA-train, Wikipedia medica IT/EN) più 22 % di replay generale. Su un test sigillato a uso singolo di 3.766 item medici a scelta multipla, a scala LoRA 0,5, l’accuratezza passa dal 61,34 % al 63,67 % (Δ = +2,34 pp, IC 95 % [+1,14; +3,53], McNemar esatto p = 1,4·10-4). MedQA e MMLU-med non sono singolarmente significativi; la guardia ARC italiana non peggiora. Limiti dichiarati: checkpoint 443 (non l’886 finale); solo log-probabilità della lettera; GSM8K e IFEval non ancora misurati; audit di rete solo statico. Non è un dispositivo medico: non usare per consigli medici. Files: article_en.pdf, article_it.pdf, code.zip (statistics, figures, LaTeX sources, original training/evaluation scripts, checksums). Adapter weights: Hugging Face odelpa/qwen3-8b-med6m; code: GitHub usergdp136/qwen3-8b-med6m. Licences: article, figures and adapter CC BY-SA 4.0; code Apache-2.0; base model Apache-2.0.

Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
med6m: a LoRA adapter for Qwen3-8B on medical multiple-choice QA — sealed-test evidence for checkpoint 443 at scale 0.5 / med6m: un adattatore LoRA per Qwen3-8B per QA medico a scelta multipla — evidenza su test sigillato — Giovanni Del Papa · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS