Synthetic Medical Vision · OpenMed
synthvision-annotated-kimi
A synthetic medical imaging dataset annotated by Kimi K2.5 for visual question answering and diagnostic evaluation.
What synthvision-annotated-kimi is
synthvision-annotated-kimi is a synthetic medical imaging dataset created by OpenMed. It forms the second half of the annotations produced within the SynthVision pipeline, serving as a companion collection to a sister set annotated by Qwen.
The dataset addresses the need for structured, reasoned annotations in medical vision tasks, specifically supporting cross-annotator validation across distinct model architectures. It is intended for computational researchers and clinical AI developers training or evaluating vision-language models on clinical imagery.
The resource contains 59,539 annotated records across medicine and biology. Each entry pairs medical image paths with multi-turn dialogues, structured observations, detailed clinical narratives, diagnostic reasoning steps, and difficulty ratings.
How it was made
Annotations were generated by the Kimi K2.5 model (a 1-trillion-parameter mixture-of-experts model with 32 billion active parameters) through Doubleword batch inference.
Each record includes a unique identifier, a relative image path, and a multi-turn clinical dialogue in ShareGPT format containing 5 to 9 turns. It also provides a 4 to 6 sentence narrative clinical report, structured key-value findings, numbered diagnostic reasoning steps, and a difficulty label classified as easy, intermediate, or hard.
Key facts
- Record count
- 59,539
- Annotator model
- Kimi K2.5 (1T MoE, 32B active)
- Batch inference platform
- Doubleword
- Primary task
- Visual question answering
- Data format
- Parquet
- Fields
- Medicine, biology
- Dialogue format
- ShareGPT (5–9 turns)
What you can use it for
- Medical visual question answering
- Fine-tuning and benchmarking multimodal models using multi-turn clinical question-and-answer exchanges grounded in medical images.
- Cross-annotator consistency validation
- Comparing diagnostic annotations, reasoning steps, and clinical reports against the sister Qwen-annotated dataset to study model-specific biases.
- Clinical diagnostic reasoning analysis
- Training or evaluating models on step-by-step diagnostic reasoning using the numbered diagnostic chains provided for each synthetic image.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from OpenMed/synthvision-annotated-kimi at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 59,539 rows of default/train
annotation_model
string_label- moonshotai/Kimi-K2.5100%
conversations
string_text- min
- 2
- median
- 3,220
- mean
- 3259.21
- max
- 7,508
difficulty
string_label- intermediate92%
- easy5%
- hard3%
id
string_text- min
- 6
- median
- 10
- mean
- 12.19
- max
- 23
image_path
string_text- min
- 75
- median
- 79
- mean
- 83.38
- max
- 95
modality
string_label- radiology88%
- pathology11%
- ct_scan1%
reasoning_chain
string_text- min
- 0
- median
- 686
- mean
- 694.78
- max
- 1,460
report
string_text- min
- 0
- median
- 814
- mean
- 820.02
- max
- 1,622
source
string_label- roco55%
- multicare42%
- pathvqa3%
structured_findings
string_text- min
- 2
- median
- 406
- mean
- 414.04
- max
- 1,479
Strengths and limits
Good at
- Provides multi-faceted annotation for each image, encompassing dialogue, narrative reports, structured findings, and numbered reasoning chains.
- Stratifies cases by difficulty rating across easy, intermediate, and hard tiers.
- Matches the schema of the Qwen-annotated split to enable cross-annotator evaluation within the SynthVision pipeline.
Watch out for
- Annotations are synthetically generated by an AI model rather than human clinicians.
- Records provide relative image paths rather than raw pixel data embedded directly in the text records.
Questions
Can I look inside synthvision-annotated-kimi without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is synthvision-annotated-kimi under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What model generated the annotations in this dataset?
The annotations were generated by Kimi K2.5, a mixture-of-experts model with 1 trillion total parameters and 32 billion active parameters, executed via Doubleword batch inference.
What information is included in each dataset entry?
Each record includes a unique identifier, relative image path, multi-turn clinical conversations in ShareGPT format, a clinical narrative report of 4 to 6 sentences, structured findings, numbered diagnostic reasoning steps, and a difficulty label.
How does this relate to other SynthVision datasets?
It serves as the second-half counterpart to the Qwen-annotated dataset in the SynthVision pipeline, designed to enable cross-annotator validation.
Written from synthvision-annotated-kimi's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.