Merlin: a report-generation reproduction plan
Separate library setup, checkpoint downloads and model-generated text for one public sample. An original research template, not a Merlin agent skill.
The complete kit in one ZIP
One ZIP: skills, instructions, installer and a public demo runner.
When to use it
Check requirements for the selected Merlin task
Plan the path from DataLoader to raw report
Plan a review of omissions and unsupported statements
What you’ll get
A one-task plan with requirements, commands, a raw-output record and review questions.
For example
Plan a Merlin(RadiologyReport=True) pilot on a public sample. Check GPU, free disk space and the official DataLoader; do not download weights or run the model.
Before you start
An original kossolapov.com template for planning Merlin report generation. Merlin is a research model, not an agent skill.
Select one task: this template covers report generation, not embeddings or segmentation. Provide GPU/VRAM, disk space and a public sample.
Check the pinned README, pyproject, demo and data terms. Code is MIT; dataset access has separate terms.
Account for approximately 25 GB of weights. Documentation reports an author test on a 48 GB A6000, not a minimum we have verified.
Trace the official DataLoader, working directory and cache. Prepare commands without downloading weights or running the model.
Keep raw text separate from paraphrases. Plan a review of omissions and unsupported claims; Merlin was not run in this release.