cs552-the-expendables/patientagent-eval-results

PatientAgent evaluation results

Public conversations, original full-dialogue G-Eval, continuity-filtered G-Eval, source-versus-generated discriminator artifacts, and compact SFT training analyses for the current PatientAgent comparison.

Current verified evaluation population

All response sources use the same source-eligible dialogue IDs from both official MTS-Dialog test splits. Eligibility is defined before generation: the ground-truth source dialogue must contain at least one patient turn.

Split	Candidates	Eligible and G-Eval scored	Excluded
Test1	200	174	26
Test2	200	178	22
Combined	400	352	48

Test1 excludes 16 doctor-family dialogues, 9 doctor-guest-clinician dialogues, and 1 mixed family/guest-clinician dialogue. Test2 excludes 11 doctor-family dialogues, 9 doctor-guest-clinician dialogues, 1 mixed family/guest-clinician dialogue, and 1 doctor-only dialogue. None contains a patient turn for the model to replace.

Original full-dialogue G-Eval results

Each value below is a dialogue-level micro-average over the same 352 eligible IDs. Split-specific results remain available for auditing but are not reported as separate headline comparisons. Hallucination and irrelevance are lower-is-better; anthropomorphism is higher-is-better.

Response source	Hallucination	Irrelevance	Anthropomorphism	Mean patient-turn chars.
Base Qwen3.5-4B	3.07	5.96	5.46	286.0
AgentClinic prompt control (Qwen3.5-4B)	4.22	5.24	7.63	171.8
SFT16	0.70	1.36	8.56	28.5
CFT-augmented SFT16 (experimental)	0.72	1.22	8.67	27.5
SFT128	0.77	1.59	8.54	28.7
KTO-SFT16	2.70	4.24	8.29	111.9
KTO-SFT128	2.26	4.08	7.65	107.2
DPO-SFT16	1.16	2.53	8.63	49.8
DPO-SFT128	1.02	2.02	8.51	38.8
Gold responses	0.30	2.04	8.95	45.7
Liu et al. released Qwen2.5-72B adapter (external system)	2.09	3.69	5.39	103.6

These are the original immutable scores over each complete generated/source dialogue. They remain published for provenance and reproducibility.

Continuity-filtered G-Eval results

A separately derived layer checks the next doctor response after every generated patient reply. When that fixed source-doctor turn is a human-confirmed non-sequitur, the derived dialogue retains the preceding patient reply and excludes the bad doctor turn and all later turns. No original dialogue or score is deleted or edited. Dialogues without a confirmed break reuse their original score dictionary byte-for-byte; only affected prefixes are scored again with the same judge and prompt.

Response source	Confirmed breaks	Hallucination	Irrelevance	Anthropomorphism	Mean patient-turn chars.
Base Qwen3.5-4B	1	3.0455	5.9574	5.4602	286.0413
AgentClinic prompt control	4	4.2074	5.2358	7.6392	171.4690
SFT16	7	0.7017	1.3466	8.5824	28.5580
CFT-augmented SFT16	7	0.7216	1.2188	8.6619	27.3850
SFT128	10	0.7472	1.5625	8.5710	28.6058
KTO-SFT16	2	2.6932	4.2358	8.3011	111.6344
KTO-SFT128	1	2.2614	4.0824	7.6534	107.2049
DPO-SFT16	2	1.1506	2.5483	8.6307	49.7885
DPO-SFT128	4	0.9915	2.0341	8.5227	38.7738
Gold responses	0	0.3011	2.0426	8.9545	45.7024
Liu Qwen2.5-72B external system	3	2.0568	3.6818	5.3864	103.4194

The final catalog covers 3,872 source-task dialogue instances. It contains 41 human-confirmed first breaks, 3,831 exact original-score reuses, 41 successful same-judge prefix scores, and 235 excluded later patient turns. Every classifier majority-positive decision and one-of-three split negative was manually reviewed, including suffixes exposed by human-overturned early stops. Across all instances, filtering changes pooled hallucination by -0.0111, irrelevance by -0.0023, anthropomorphism by +0.0067, and mean response length by -0.0850 characters; aggregate conclusions are materially unchanged.

The AgentClinic row isolates prompting from model choice: it uses the same no-adapter Qwen3.5-4B backend, case facts, dialogue-history representation, generation settings, eligible IDs, and judge as the base row. Its system message is the patient prompt published verbatim in AgentClinic Appendix L.2, with only the published target language and Symptoms Information placeholders filled (English and the current case facts, respectively). It is a prompt-only control, not a reproduction of AgentClinic's full heterogeneous system.

The Liu et al. row is an external-system reference, not a controlled method comparison. It uses the authors' released rank-32 LoRA at revision 89520a20ebfce5752197404ebb65892824024728 on its declared Qwen2.5-72B-Instruct base at revision 495f39366efef23836d0cfae4fbe635880d2be31, loaded exactly in BF16 on two H200 GPUs. The release was trained on Chinese synthetic dialogues and was evaluated here without output cleaning on English MTS-Dialog. Model family, parameter count, training data, and language therefore all differ from the Qwen3.5-4B rows. The release does not include the training data, dialogue-strategy flows, or training code needed for a same-backend reproduction.

Current source-versus-generated discriminator

Gold supplies the MTS-Dialog source reference and provenance label, not a scored model. The full eligible population contains 352 dialogues and 1,545 patient turns. A 25-call throughput pilot projected the full two-order protocol at 96.0 hours (115.2 hours with a 20% buffer), so seed 20260814 fixed the largest conservative split-stratified whole-dialogue sample that fit the approved runtime budget before outcomes were inspected: 53 dialogues and 280 patient turns (Test1: 26/130; Test2: 27/150). All eight learned response sources use this identical manifest and both presentation orders.

Response source	Detection micro (95% cluster CI)	Dialogue macro (95% cluster CI)	Fooling rate	Exact ties
Base Qwen3.5-4B	0.698 [0.624, 0.752]	0.626 [0.539, 0.713]	0.302	0
SFT16	0.277 [0.221, 0.354]	0.329 [0.260, 0.398]	0.723	24
SFT128	0.318 [0.281, 0.363]	0.353 [0.286, 0.416]	0.682	25
CFT-augmented SFT16	0.341 [0.294, 0.412]	0.397 [0.321, 0.480]	0.659	21
KTO-SFT16	0.809 [0.740, 0.864]	0.765 [0.688, 0.836]	0.191	0
KTO-SFT128	0.755 [0.664, 0.816]	0.679 [0.592, 0.760]	0.245	1
DPO-SFT16	0.489 [0.423, 0.560]	0.509 [0.426, 0.591]	0.511	11
DPO-SFT128	0.388 [0.332, 0.463]	0.396 [0.322, 0.470]	0.613	13

Detection is the judge's accuracy at locating the learned response; fooling is its complement. Non-identical pairs are shown in both A/B orders, while text-identical pairs receive an analytical 0.5 without an API call. Confidence intervals use a split-stratified dialogue-cluster bootstrap. All 4,290 judge calls succeeded on the first attempt, and an independent publication validator reconstructed the manifests and task plan, checked every checkpoint record and raw label, and recomputed every metric and interval.

Manual inspection found a substantial length/style confound. The judge often treats a terse generated answer as source text and a longer source answer as generated; highly detectable base and KTO cases include verbosity, truncation, repetition, unsupported detail, and non-answers. Detection is 15.0--28.2 percentage points higher when the generated response occupies slot B. CFT averaged 30.8 characters per turn versus 45.8 for the references and had a 21.4-point slot-B advantage. Its detection accuracy was 6.4 points above SFT16, but the paired split-stratified dialogue-cluster interval included zero [-0.7, 11.9]. The counterbalanced result therefore measures source-style detectability, not human-likeness or clinical quality.

Artifact layout

• current/test1/<variant>/conversations.json and results.json: generated or gold conversations plus complete 174-dialogue G-Eval scores.
• current/test2/<variant>/conversations.json and results.json: generated or gold conversations plus complete 178-dialogue G-Eval scores.
• current/combined/<variant>/results.json: validated 352-dialogue G-Eval micro-average with split-tagged per-dialogue records.
• current/continuity-filtered-geval-v1/: derived 22-task filtered score catalog, full/suffix continuity ledgers, completed human adjudications, independent validations, metric deltas, and a SHA-256 publication manifest.
• current/training-analysis/sft-current-v1/: validated six-run rank/learning-rate sweep and two-seed paired follow-up summaries and curves; adapters are not duplicated in this dataset.
• current/discriminator/manifest.json: frozen common 53-dialogue/280-turn discriminator manifest.
• current/discriminator/results.json: pair-level decisions, aggregates, strata, and bootstrap intervals.
• current/discriminator/full_checkpoint.sqlite3: fingerprint-bound transactional raw-call checkpoint.
• current/discriminator/validation.json: independent recomputation and publication-gate report.
• current/discriminator/manual_audit_samples.json: representative and anomalous examples selected for manual review.

G-Eval variants are base, agentclinic-qwen35, sft16, sft128, cft-r16 (experimental CFT-augmented SFT16), kto-sft16, kto-sft128, dpo-sft16, dpo-sft128, gold, and liu-qwen25-72b-bf16 (heterogeneous external system). Every split result has zero missing IDs and records the SHA-256 fingerprint of its exact eligible conversation cache.

Interpretation limits

Scores come from one LLM judge family. G-Eval uses one Kimi-K2.7-Code pass with a 100,000-token output ceiling; exact duplicate conversations can receive slightly different scores on independent calls, and extracted case facts occasionally omit details present in the source dialogue. Manual review of 12 representative or extreme CFT dialogues found 11 explanations broadly aligned with the rubric and one clear patient/non-patient role-attribution error in a mixed child/parent dialogue; no post-hoc score was changed. A separate deterministic audit of 12 representative or extreme Liu dialogues found 10 explanations broadly aligned, one anthropomorphism explanation that missed a trailing bracketed control tag, and one irrelevance explanation that incorrectly treated a symptom report as off-topic after the doctor asked how the patient felt; again, no score was changed. The facts-only judge can therefore penalize a patient for repeating a detail introduced by the clinician in the visible history. The literal AgentClinic prompt also assumes symptom-centric input and instructs the patient not to reveal the disease, whereas the current case facts can contain diagnoses, medication, surgery, allergy, demographics, family history, and social history; this is an intrinsic compatibility limitation of the exact-prompt control. The Liu release produced leading bracketed act/affect annotations in 1,467 of 1,545 patient turns, substantive CJK text in 232 turns, and literal [STOP] markers in 47 turns; raw outputs were intentionally preserved. Discriminator confidence intervals quantify dialogue sampling, not judge variability, and its runtime-limited dialogue sample covers 53 of 352 eligible dialogues. These artifacts measure fact-grounded automated judgments and source-style detectability, not clinical quality or a perfect gold ceiling. Human validation remains absent; the Liu external-system comparison and downstream-clinician benchmark are reported with their respective confounds.

Historical artifacts

The older 183-dialogue snapshot remains available in MTS-Dataset-preprocessed. It is separate and must not be mixed with this evaluation.

Intended use

Research benchmark only. These generated dialogues can hallucinate and are not medical advice, a medical device, or evidence that a doctor model is clinically safe.