Synthetic Medical Vision · OpenMed
synthvision-annotated-qwen
A synthetic medical imaging dataset annotated by Qwen 3.5 for training multimodal clinical models.
What synthvision-annotated-qwen is
synthvision-annotated-qwen is a medical imaging dataset developed by OpenMed for training and validating multimodal artificial intelligence models in medicine and biology. It provides annotations for synthetic medical imagery as part of OpenMed's SynthVision pipeline, supporting tasks such as visual question answering.
The dataset contains 59,476 records generated during the first half of the SynthVision annotation process. Each record associates a synthetic medical image with rich clinical data, including multi-turn clinical dialogue, structured findings, reasoning steps, and narrative reports.
How it was made
Annotations were generated across 59,476 medical images using Qwen 3.5, a 397-billion parameter mixture-of-experts (MoE) language model with 17 billion active parameters, executed via Doubleword batch inference.
Each entry follows a structured schema: a unique identifier, relative image path, multi-turn clinical conversations in ShareGPT format (5 to 9 turns), a 4-to-6 sentence clinical narrative report, a dictionary mapping clinical finding names to values, a numbered diagnostic reasoning chain, and a difficulty rating categorized as easy, intermediate, or hard.
Key facts
- Record count
- 59,476
- Annotation model
- Qwen 3.5 (397B MoE, 17B active)
- Inference platform
- Doubleword
- Data format
- Parquet
- Task category
- Visual question answering
- Fields
- Medicine, biology
What you can use it for
- Medical Visual Question Answering
- Training and evaluating multimodal vision-language models on clinical question answering using multi-turn conversational exchanges.
- Clinical Diagnostic Reasoning
- Teaching multimodal models step-by-step diagnostic workflows using the provided numbered reasoning chains.
- Automated Medical Report Generation
- Benchmarking models on synthesizing short clinical narrative reports and structured findings directly from medical images.
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-qwen at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 59,476 rows of default/train
annotation_model
string_label- Qwen/Qwen3.5-397B-A17B100%
conversations
string_text- min
- 2
- median
- 1,800
- mean
- 1819.20
- max
- 3,649
difficulty
string_label- intermediate91%
- easy6%
- hard3%
id
string_text- min
- 6
- median
- 10
- mean
- 12.20
- max
- 23
image_path
string_text- min
- 75
- median
- 79
- mean
- 83.40
- max
- 95
modality
string_label- radiology88%
- pathology12%
- ct_scan1%
reasoning_chain
string_text- min
- 241
- median
- 586
- mean
- 593.03
- max
- 1,168
report
string_text- min
- 273
- median
- 737
- mean
- 737.64
- max
- 1,321
source
string_label- roco55%
- multicare42%
- pathvqa3%
structured_findings
string_text- min
- 2
- median
- 220
- mean
- 224.52
- max
- 755
Strengths and limits
Good at
- Provides multi-faceted annotations including multi-turn dialogue, narrative reports, structured findings, and step-by-step reasoning for every image.
- Categorises records by diagnostic difficulty into easy, intermediate, and hard tiers.
- Generated using a large-scale mixture-of-experts model (Qwen 3.5 with 397B parameters).
Watch out for
- Represents only the first-half annotations of the full SynthVision pipeline.
- Contains synthetic rather than real-world patient medical images.
Questions
Can I look inside synthvision-annotated-qwen 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-qwen under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What information does each record contain?
Each record includes a unique ID, a relative image path, a multi-turn conversation in ShareGPT format (5 to 9 turns), a 4-to-6 sentence clinical report, structured findings as key-value pairs, a numbered reasoning chain, and a difficulty level.
How were the annotations created?
They were produced by running batch inference with the Qwen 3.5 MoE model (397B total, 17B active parameters) through Doubleword.
How are the conversation turns structured?
The dialogue consists of 5 to 9 turns per image, formatted in standard multi-turn ShareGPT format.
Written from synthvision-annotated-qwen'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.