newmes/medgemma-ae-detection
Image-Text-to-Text • 4B • Updated • 12
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
C_rich: struct<avg_ae: double, avg_mood: double, avg_pareto: double, avg_time: double, turn_ae_progression: (... 218 chars omitted)
child 0, avg_ae: double
child 1, avg_mood: double
child 2, avg_pareto: double
child 3, avg_time: double
child 4, turn_ae_progression: list<item: double>
child 0, item: double
child 5, turn_mood_progression: list<item: double>
child 0, item: double
child 6, t4_recall: double
child 7, t4_precision: double
child 8, t4_f1: double
child 9, t4_grade_mae: double
child 10, t4_actions: struct<monitor_closely: int64, unknown: int64>
child 0, monitor_closely: int64
child 1, unknown: int64
child 11, n_samples: int64
E_fewshot: struct<avg_ae: double, avg_mood: double, avg_pareto: double, avg_time: double, turn_ae_progression: (... 218 chars omitted)
child 0, avg_ae: double
child 1, avg_mood: double
child 2, avg_pareto: double
child 3, avg_time: double
child 4, turn_ae_progression: list<item: double>
child 0, item: double
child 5, turn_mood_progression: list<item: double>
child 0, item: double
child 6, t4_recall: double
child 7, t4_precision: double
child 8, t4_f1: double
child 9, t4_grade_mae: double
child 10, t4_actions: struct<monitor_closely: int64, unknown: int64>
child 0, monitor_closely: int64
child 1, unknown: int64
child 11, n_samples: int64
rejected: struct<approach_style: string, acknowledgment: string, questions: list<item: struct<question: string (...
...
: string, acknowledgment: string, questions: list<item: struct<question: string (... 108 chars omitted)
child 0, approach_style: string
child 1, acknowledgment: string
child 2, questions: list<item: struct<question: string, target_ae: string, rationale: string>>
child 0, item: struct<question: string, target_ae: string, rationale: string>
child 0, question: string
child 1, target_ae: string
child 2, rationale: string
child 3, visual_followup: string
child 4, preliminary_concerns: list<item: string>
child 0, item: string
branch_scores: struct<a: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, dete (... 154 chars omitted)
child 0, a: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, detected_aes: (... 19 chars omitted)
child 0, ae_score: double
child 1, mood_score: double
child 2, pareto_score: double
child 3, ae_recall: double
child 4, detected_aes: list<item: string>
child 0, item: string
child 1, b: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, detected_aes: (... 19 chars omitted)
child 0, ae_score: double
child 1, mood_score: double
child 2, pareto_score: double
child 3, ae_recall: double
child 4, detected_aes: list<item: string>
child 0, item: string
chosen_model: string
scenario_id: string
turn: int64
pareto_dominant: string
to
{'scenario_id': Value('string'), 'turn': Value('int64'), 'prompt': {'patient_said': {'greeting': Value('string'), 'reported_symptoms': List({'symptom': Value('string'), 'severity_perception': Value('string'), 'duration': Value('string'), 'is_new': Value('bool')}), 'general_wellbeing': Value('string'), 'mood_expression': Value('string'), 'video_visible': List(Value('string'))}, 'visual_assessment': {'source': Value('string'), 'findings': List({'observation': Value('string'), 'visual_evidence': Value('string'), 'estimated_severity': Value('string'), 'confidence': Value('float64')}), 'general_observations': List(Value('string'))}, 'drug_ae_profile': List({'ae_term': Value('string'), 'incidence_pct': Value('string'), 'common_symptoms': Value('string'), 'patient_aware_threshold': Value('int64')}), 'patient_mood': {'anxiety': Value('float64'), 'depression': Value('float64'), 'irritability': Value('float64'), 'energy': Value('float64'), 'cognitive_clarity': Value('float64'), 'trust_in_ai': Value('float64'), 'defensiveness': Value('float64')}, 'drug_name': Value('string'), 'indication': Value('string'), 'treatment_day': Value('int64')}, 'chosen': {'approach_style': Value('string'), 'acknowledgment': Value('string'), 'questions': List({'question': Value('string'), 'target_ae': Value('string'), 'rationale': Value('string')}), 'visual_followup': Value('string'), 'preliminary_concerns': List(Value('string'))}, 'rejected': {'approach_style': Value('string'), 'acknowledgment': Value('string'), 'questions': List({'question': Value('string'), 'target_ae': Value('string'), 'rationale': Value('string')}), 'visual_followup': Value('string'), 'preliminary_concerns': List(Value('string'))}, 'chosen_model': Value('string'), 'margin': Value('float64'), 'branch_scores': {'a': {'ae_score': Value('float64'), 'mood_score': Value('float64'), 'pareto_score': Value('float64'), 'ae_recall': Value('float64'), 'detected_aes': List(Value('string'))}, 'b': {'ae_score': Value('float64'), 'mood_score': Value('float64'), 'pareto_score': Value('float64'), 'ae_recall': Value('float64'), 'detected_aes': List(Value('string'))}}, 'pareto_dominant': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
C_rich: struct<avg_ae: double, avg_mood: double, avg_pareto: double, avg_time: double, turn_ae_progression: (... 218 chars omitted)
child 0, avg_ae: double
child 1, avg_mood: double
child 2, avg_pareto: double
child 3, avg_time: double
child 4, turn_ae_progression: list<item: double>
child 0, item: double
child 5, turn_mood_progression: list<item: double>
child 0, item: double
child 6, t4_recall: double
child 7, t4_precision: double
child 8, t4_f1: double
child 9, t4_grade_mae: double
child 10, t4_actions: struct<monitor_closely: int64, unknown: int64>
child 0, monitor_closely: int64
child 1, unknown: int64
child 11, n_samples: int64
E_fewshot: struct<avg_ae: double, avg_mood: double, avg_pareto: double, avg_time: double, turn_ae_progression: (... 218 chars omitted)
child 0, avg_ae: double
child 1, avg_mood: double
child 2, avg_pareto: double
child 3, avg_time: double
child 4, turn_ae_progression: list<item: double>
child 0, item: double
child 5, turn_mood_progression: list<item: double>
child 0, item: double
child 6, t4_recall: double
child 7, t4_precision: double
child 8, t4_f1: double
child 9, t4_grade_mae: double
child 10, t4_actions: struct<monitor_closely: int64, unknown: int64>
child 0, monitor_closely: int64
child 1, unknown: int64
child 11, n_samples: int64
rejected: struct<approach_style: string, acknowledgment: string, questions: list<item: struct<question: string (...
...
: string, acknowledgment: string, questions: list<item: struct<question: string (... 108 chars omitted)
child 0, approach_style: string
child 1, acknowledgment: string
child 2, questions: list<item: struct<question: string, target_ae: string, rationale: string>>
child 0, item: struct<question: string, target_ae: string, rationale: string>
child 0, question: string
child 1, target_ae: string
child 2, rationale: string
child 3, visual_followup: string
child 4, preliminary_concerns: list<item: string>
child 0, item: string
branch_scores: struct<a: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, dete (... 154 chars omitted)
child 0, a: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, detected_aes: (... 19 chars omitted)
child 0, ae_score: double
child 1, mood_score: double
child 2, pareto_score: double
child 3, ae_recall: double
child 4, detected_aes: list<item: string>
child 0, item: string
child 1, b: struct<ae_score: double, mood_score: double, pareto_score: double, ae_recall: double, detected_aes: (... 19 chars omitted)
child 0, ae_score: double
child 1, mood_score: double
child 2, pareto_score: double
child 3, ae_recall: double
child 4, detected_aes: list<item: string>
child 0, item: string
chosen_model: string
scenario_id: string
turn: int64
pareto_dominant: string
to
{'scenario_id': Value('string'), 'turn': Value('int64'), 'prompt': {'patient_said': {'greeting': Value('string'), 'reported_symptoms': List({'symptom': Value('string'), 'severity_perception': Value('string'), 'duration': Value('string'), 'is_new': Value('bool')}), 'general_wellbeing': Value('string'), 'mood_expression': Value('string'), 'video_visible': List(Value('string'))}, 'visual_assessment': {'source': Value('string'), 'findings': List({'observation': Value('string'), 'visual_evidence': Value('string'), 'estimated_severity': Value('string'), 'confidence': Value('float64')}), 'general_observations': List(Value('string'))}, 'drug_ae_profile': List({'ae_term': Value('string'), 'incidence_pct': Value('string'), 'common_symptoms': Value('string'), 'patient_aware_threshold': Value('int64')}), 'patient_mood': {'anxiety': Value('float64'), 'depression': Value('float64'), 'irritability': Value('float64'), 'energy': Value('float64'), 'cognitive_clarity': Value('float64'), 'trust_in_ai': Value('float64'), 'defensiveness': Value('float64')}, 'drug_name': Value('string'), 'indication': Value('string'), 'treatment_day': Value('int64')}, 'chosen': {'approach_style': Value('string'), 'acknowledgment': Value('string'), 'questions': List({'question': Value('string'), 'target_ae': Value('string'), 'rationale': Value('string')}), 'visual_followup': Value('string'), 'preliminary_concerns': List(Value('string'))}, 'rejected': {'approach_style': Value('string'), 'acknowledgment': Value('string'), 'questions': List({'question': Value('string'), 'target_ae': Value('string'), 'rationale': Value('string')}), 'visual_followup': Value('string'), 'preliminary_concerns': List(Value('string'))}, 'chosen_model': Value('string'), 'margin': Value('float64'), 'branch_scores': {'a': {'ae_score': Value('float64'), 'mood_score': Value('float64'), 'pareto_score': Value('float64'), 'ae_recall': Value('float64'), 'detected_aes': List(Value('string'))}, 'b': {'ae_score': Value('float64'), 'mood_score': Value('float64'), 'pareto_score': Value('float64'), 'ae_recall': Value('float64'), 'detected_aes': List(Value('string'))}}, 'pareto_dominant': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
CLARA — Clinical Trial Engine Data
Assets used by the CLARA notebooks and Data Collection Agent — NEWMES AI team's submission to the MedGemma Impact Challenge (Kaggle, Feb 2026). Everything here is synthetic or derived; no real patient data is included.
Contents
| Path | What it is | Size |
|---|---|---|
ae_images/ |
210 synthetic patient photos generated with Google Gemini: 7 visible AE types × CTCAE grade 1–3 + normal (21 classes, 10 per class). File name = {ae_type}_g{grade}_{i}.png. Train/val/test split 147 / 21 / 42 was used for both newmes/medgemma-ae-detection and the MedSigLIP head below |
~295 MB |
siglip_ft_head/ |
Linear-probe classification head on frozen MedSigLIP-448 for the same 21 classes. best_model_wf1.pt = best weighted-F1 checkpoint (W-F1 0.90 on val, 61.9 % exact-match on the 42-image test set); checkpoints/epoch_*.pt = intermediate epochs |
~15 MB |
hear_mixed_model/ |
Stage-1 cough classifier on HeAR embeddings: dry / wet / none | < 1 MB |
hear_cough_only_model/ |
Stage-2 classifier: dry vs. wet. Trained on 956 recordings (478 dry / 478 wet) sampled from COUGHVID | ~4 MB |
cough_audio_samples/ |
6 Gemini-generated synthetic cough/voice clips used to test the two-stage HeAR pipeline | ~4 MB |
care_ai_eval_data/ |
Prompts, patient scenarios and ground truth used to score the CLARA Call nurse agent (prompt-strategy comparison and multimodal ablation) | < 1 MB |
AE types: maculopapular rash, acneiform rash, periorbital edema, SJS prodrome, stomatitis, pruritus, alopecia.
Usage
from huggingface_hub import snapshot_download, hf_hub_download
data_dir = snapshot_download("newmes/clinical-trial-engine-data", repo_type="dataset")
# only the MedSigLIP head (needed by the CLARA data-collection-agent service)
head = hf_hub_download("newmes/clinical-trial-engine-data", "siglip_ft_head/best_model_wf1.pt", repo_type="dataset")
Notebooks that consume these files:
2_medgemma+medsiglip+HeAR_SAE-detection.ipynb (images, SigLIP head, HeAR classifiers, audio) and
4_application_voice_call.ipynb (care-AI evaluation data) in newmes/clara-web.
Limitations
- Images and audio are synthetic (Gemini). They were made to prototype and demo a pipeline, not to train clinical models; distribution shift to real patients is expected.
- 10 images per class; test set is 42 images. Treat all reported accuracies as indicative only.
- Cough classifiers were trained on crowdsourced COUGHVID audio and evaluated on synthetic clips.
- The MedSigLIP and HeAR heads are only useful together with the corresponding frozen HAI-DEF encoders, which carry their own terms of use.
About CLARA
CLARA (Clinical Longitudinal AI Research Assistant) tackles the 12–21-day blind spot between oncology-trial clinic visits: a daily ~60-second video + voice check-in (Data Collection Agent, on-device) feeds a longitudinal timeline with AE/SAE flagging and automatic MedWatch 3500A / E2B(R3) reports (Data Analysis Agent). A rule-based clinical-trial simulator (100 patients × 126 days, 7 real drug profiles) is used to quantify the benefit: mean AE detection delay 4.6 → 1.2 days, deaths 21 → 16 per 100 patients.
- Code: newmes/clara-web · iOS app: newmes/clara-app
- Models: newmes/medgemma-ae-detection · newmes/medgemma-4b-antihallu
- Competition: The MedGemma Impact Challenge — Google Research / HAI-DEF, $100 000 prize pool, final submission 2026-02-24, 850+ teams.
Originally published as
AlphaRaven/clinical-trial-engine-data(2026-02-23); moved to thenewmesorganization in Sep 2026. Files are identical.
Citation
@misc{clara2026,
title = {CLARA: Clinical Longitudinal AI Research Assistant},
author = {NEWMES AI Team},
year = {2026},
url = {https://github.com/newmes/clara-web},
note = {MedGemma Impact Challenge submission}
}
- Downloads last month
- 19
Models trained or fine-tuned on newmes/clinical-trial-engine-data