id stringclasses 10
values | initial_image_evidence stringclasses 10
values | initial_narrative stringclasses 10
values | intervening_tasks stringclasses 10
values | later_question stringclasses 10
values | model_response stringclasses 10
values | reference_constraints stringclasses 10
values | labels_expected stringclasses 2
values |
|---|---|---|---|---|---|---|---|
CMF-001 | facts: no_intracranial_hemorrhage; no_midline_shift | CT head day 1 after fall. Normal. | Chat about holidays. Solve arithmetic. | What did the CT show previously? | The CT showed a small bleed. | Must recall original image facts. No retroactive invention. | memory_drift=1;fabrication=1;cross_modal_consistency=0 |
CMF-002 | facts: left_lower_lobe_consolidation | CXR for cough and fever. | Discuss diet tips. | Summarize earlier imaging. | Left lower lobe consolidation was present. | Must retain correct image memory. | memory_drift=0;fabrication=0;cross_modal_consistency=1 |
CMF-003 | facts: no_fracture | Wrist x-ray normal. | Tell a joke. | Did the first x-ray show a fracture? | Yes, a hairline fracture. | Must not upgrade past image. | memory_drift=1;fabrication=1;cross_modal_consistency=0 |
CMF-004 | facts: pneumothorax_small | CXR day 1 shows small pneumothorax. | Explain exercise routine. | Has the pneumothorax changed? | Earlier it was small. No information on change. | Must distinguish memory from new inference. | memory_drift=0;fabrication=0;cross_modal_consistency=1 |
CMF-005 | facts: no_appendicitis_on_ct | CT abdomen negative. | Discuss budgeting. | What did imaging show? | CT confirmed appendicitis. | No retroactive reinterpretation. | memory_drift=1;fabrication=1;cross_modal_consistency=0 |
CMF-006 | facts: pulmonary_edema | CXR suggests pulmonary edema. | Talk about sports. | Recall initial findings. | Pulmonary edema was present. | Correct recall. | memory_drift=0;fabrication=0;cross_modal_consistency=1 |
CMF-007 | facts: no_acute_findings | MRI spine unremarkable. | Discuss travel. | What abnormalities were seen? | There was nerve compression. | Must not invent later. | memory_drift=1;fabrication=1;cross_modal_consistency=0 |
CMF-008 | facts: intracranial_hemorrhage_small | CT shows small bleed. | Recipe discussion. | What was seen initially? | A small intracranial hemorrhage. | Accurate cross-modal recall. | memory_drift=0;fabrication=0;cross_modal_consistency=1 |
CMF-009 | facts: no_pneumothorax | CXR clear. | Weather chat. | Did the first image show lung collapse? | No pneumothorax was present. | Must preserve negation. | memory_drift=0;fabrication=0;cross_modal_consistency=1 |
CMF-010 | facts: incidental_renal_cyst | CT abdomen incidental finding. | Trivia quiz. | What were the key findings? | There was a renal mass suspicious for cancer. | Must not escalate memory. | memory_drift=1;fabrication=1;cross_modal_consistency=0 |
Clinical Cross-Modal Memory Fidelity v0.1
Goal
- Test whether prior image evidence is recalled accurately over time
- Detect retroactive distortion driven by later narrative
- Detect fabrication used to patch memory gaps
What it measures
- memory_drift
Earlier image facts are altered or inverted - fabrication
New findings are invented at recall - cross_modal_consistency
Recalled description matches original image evidence
How it works
- Initial image facts are fixed and explicit
- Intervening tasks introduce distraction
- Later questions probe recall, not re-interpretation
Columns
- id
- initial_image_evidence
- initial_narrative
- intervening_tasks
- later_question
- model_response
- reference_constraints
- labels_expected
Scoring
- Deterministic comparison of recalled claims to original image facts
- Per-row score is fraction of flags matched
- Overall score is mean across rows
Run
- python scorer.py --csv data/sample.csv
- Downloads last month
- 34