The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Copy and send to your coding agent
Please download the judge package from:
https://huggingface.co/datasets/BenluW/3dmed-error-judge100-single-agent/resolve/single-agent-v2.1-20261105/error_type_judge_SINGLE_AGENT_100_EN_v2_1_20261105.zip
Extract it, install package requirements, and run `python tools/verify_package.py` from the extracted package directory. Follow `prompts/coordinator.txt` and `docs/WORKFLOW.md` to process all 100 cases.
For EVERY case, you MUST create a NEW subagent with a distinct agent/session ID and no inherited conversation history. Never reuse a subagent across cases or perform case annotation in the coordinator. Each newly created subagent must own ONE COMPLETE case and perform every phase itself: full E1–E6/C1–C2 review, E1a–E1d subtyping, targeted E5/C1 re-review, candidate self-check, validation, and final assembly. Keep the same subagent throughout that case; do not create separate agents for its phases. Run different cases in parallel within your platform’s capacity.
Use the complete supplied evidence, prompts, and schemas. No images or new diagnostic inference are needed. Preserve uncertain/unresolved findings and report execution failures explicitly rather than silently dropping cases.
Finally, return `judge100_results.zip` containing the complete `outputs/` directory, generated per-case tasks and dependent requests, and an execution summary recording each packet ID, its newly created subagent ID, actual model used, completion status, and any failures. Include all intermediate phase outputs, validation records, and final case artifacts—not only a summary table.
3DMed Error Judge 100: one subagent per complete case
A ready-to-use, English annotation workflow for 100 complete case-response packets: 100 diagnosis answers + 257 follow-up answers from 12 multimodal models. Each subagent owns one entire case, from its first error audit through E1 subtyping, targeted E5/C1 re-review, candidate self-check and final assembly.
Public instruction revision: v2.1 for error-type-single-agent-case-workflow-en-v2.
MANDATORY: For EVERY case, CREATE A NEW subagent with a distinct agent/session ID and no inherited history. Never reuse a subagent across cases. The coordinator must not annotate cases itself. Use fork_turns="none" or the platform equivalent. If new isolated subagents cannot be created, report that limitation and leave the case unrun. The same newly created subagent completes every phase of its case.
- Download the complete ZIP (about 4.9 MB).
- Coordinator prompt: give this to the primary agent.
- Complete-case subagent prompt.
- One complete example task.
- Full operating guide.
- Package README.
Download and start
No login is required for this public repository. Download the entire repository:
python -m pip install -U huggingface_hub hf_xet
hf download BenluW/3dmed-error-judge100-single-agent \
--repo-type dataset --revision single-agent-v2.1-20261105 --local-dir ./judge100
cd judge100/package
python -m pip install -r requirements.txt
python tools/verify_package.py
Give prompts/coordinator.txt to the primary agent, with PACKAGE_ROOT set to
this package's absolute path and PACKET_IDS set to ALL or an assigned list.
The coordinator dispatches one subagent for each complete case. To export one
case's task manually:
python tools/export_worker_task.py --packet-id new_001 --output work/new_001/case.task.json
The same subagent executes all phases itself. The offline tools prepare inputs,
validate outputs and assemble results; they do not call APIs or run inference.
Return outputs/, including outputs/final/<packet_id>.json, phase outputs and
validation records. Keep execution metadata when available.
Workflow
- Full E1–E6 / C1–C2 review using source-grounded definitions and evidence gates.
- E1a–E1d subtyping of the E1 events found in that case.
- Targeted E5 / C1 re-review of the complete original evidence.
- Evidence-based self-check of the proposed E5/C1 candidates.
- Deterministic assembly with source-qualified links and unresolved judgments.
All phases of a case share the same subagent context. The targeted re-review and self-check are not independent or blinded reviews. The previous multi-agent-per-case design is a different protocol. Final artifacts identify the intended same-agent review design; execution logs are needed to establish what actually ran.
Data and interpretation
The source cohort is the fixed paper1000-20261031 set. These 100 model/case
answer groups had not received error-type judgments in the known earlier
queues. Another model's answer for the same case may already have been judged;
answer-correctness scoring may also exist. The subset is balanced by model,
not a random clinical sample.
Each packet includes the complete questions, user-visible model answers and explanations, reference/source text, evidence quotations, clinical context and image metadata. No image pixels or model weights are included or required. This is a text-only, source-grounded audit. Source silence does not establish that an image finding is absent. Preserve uncertainty, unresolved items and image-adjudication needs under the supplied rules. References are provisional and are not a newly physician-adjudicated gold standard.
The original questions, answers, source quotations, classification definitions and output schemas are preserved. No annotations were generated for this release. Tests validate workflow behavior and artifact integrity, not clinical accuracy. No existing real error-type or subtype labels are included; test fixtures are explicitly synthetic.
Repository layout
README.md This dataset card
SOURCES.csv Original case identifiers and source links
RELEASE.json Frozen release identity and archive SHA-256
error_type_judge_SINGLE_AGENT_100_EN_v2_1_20261105.zip
package/
README.md
prompts/coordinator.txt Primary-agent entry point
prompts/case_worker.txt One complete case per subagent
data/original/ 100 evidence packets
schemas/ Required JSON output contracts
requests/ Prepared inputs
examples/ Complete-case example
tools/ Offline preparation, validation, assembly
docs/WORKFLOW.md Exact per-case commands
manifest.json, provenance/ Original source lineage
workflow.json, SHA256SUMS.json Protocol and integrity metadata
Source attribution and terms
The case evidence contains material derived from Radiopaedia case records.
SOURCES.csv lists the retained source case IDs and links constructed from
those IDs; this release does not claim a fresh page-by-page source-content
verification. Source references and technical lineage are retained in each
packet and its provenance. Original third-party material remains subject to
its applicable terms; this upload does not assign it a new blanket license.
Model identifiers and answer provenance are recorded for reproducibility.
- Downloads last month
- 168