QWEN3.5-2B · FINDINGS TO IMPRESSION
Compare the reports.
See what changed.
Explore reference reports alongside untouched baselines and non-thinking fine-tuned models. Every model sees the same Findings, without a persona instruction.
Explore saved epoch-3 generations with learned persona tokens, matched unconditioned models, and the untouched base. All see the same Findings; conditioned models also receive the displayed persona token. This is not live inference.
—test reports
4model outputs / report
40 / 10ID / OOD personas
How to read this dashboard Research only
ID/OOD follows the selected test catalog: ID = persona ranks 1–40; OOD = ranks 41–50. Evaluating a model on the other dataset does not change this label here. A separate badge identifies cross-dataset models. Base-model pretraining exposure is unknown.
Conditioning and actual exposure: joint and disjoint use distinct persona catalogs. Own-dataset ranks 1–40 receive their dedicated token; held-out ranks 41–50 and every cross-dataset report receive <|persona_unknown|>. Cards separately show dataset ID/OOD and persona seen/unseen. Cross-dataset generation combines dataset OOD with unseen personas, even when the report has a catalog-ID label. Identical token strings across models do not identify the same semantic persona.
Scores: GREEN measures reference-relative clinical agreement; Style measures agreement on applicable style features; ROUGE-L measures wording overlap; Checks are deterministic reference-relative rules. All scorers compare Impressions only, not Findings. GREEN and Style show mean ± sample SD across three calls to the same judge, not a confidence interval.
Decoding: this is the original experiment: non-thinking temperature 0.7, top-p 0.8, presence penalty 1.5; thinking 0.6 / 0.95 / 1.5; top-k 20 in both. These are not Qwen's recommended text settings. The persona experiment uses a separate tab so the profiles are never mixed. Trained thinking models are excluded.
Matched comparison: all five models use non-thinking text decoding: temperature 1, top-p 1, top-k 20, presence penalty 2, min-p 0, repetition penalty 1; maximum 576 Impression tokens. Both conditioned and unconditioned SFT models are final epoch-3 checkpoints (joint step 923; disjoint step 927), not the older selected steps 300/600. Only headline generations appear here, not wrong/unknown-token control ablations. Training used fixed approximately 5% unknown-token dropout. Experimental setup and aggregate results with confidence intervals.
Public data: the owner approved these non-identifying, open-dataset-derived records for public release. Underlying source terms still apply. Raw reasoning, judge responses, source-record identifiers, and patient metadata are not published. This site is not a clinical decision tool.
Loading report catalog...
No matching reports
Try another term or reset your filters.