// Add a new pathology foundation model to the README (excluding magnification). Use when the user asks to add a model, paper, or feature extractor to the list, or mentions a new paper in the papers/ directory.
Calculate and add the effective magnification for a pathology foundation model already listed in the README. Use when the user asks to add magnification for a model that is already in the table.
Find and thoroughly read a pathology foundation model paper from the papers/ directory. Use as a prerequisite before adding model information or magnification to the README.
| name | add-model |
| description | Add a new pathology foundation model to the README (excluding magnification). Use when the user asks to add a model, paper, or feature extractor to the list, or mentions a new paper in the papers/ directory. |
Add a new pathology foundation model row to README.md by extracting information from its paper PDF and associated GitHub repository. This skill covers all fields except magnification, which is handled separately by the add-magnification skill.
First, follow the read-paper skill (.claude/skills/read-paper/SKILL.md) to find and thoroughly read the paper. Then return here to add the model row.
The README has two tables:
| Table | Criteria |
|---|---|
| Patch-level models | Produces patch/tile embeddings (most models) |
| Slide-level / patient-level models | Produces WSI-level or patient-level embeddings without supervision |
For patch-level models, extract these fields:
| Field | Description | Notes |
|---|---|---|
| Name | Model name with paper link | Use **bold** if >100K WSIs |
| Group | Research group/institution | Link to lab website if available |
| Weights | :white_check_mark: or :x: | Check GitHub/HuggingFace |
| Released | Date + link to first release | Format: Mon YYYY[*](link) |
| SSL | Self-supervised learning method | Link to paper if novel method |
| WSIs | Number of whole-slide images | Use **bold** if >100K; round to 2 sig figs |
| Tiles | Number of patches/tiles | Round to 2 sig figs |
| Patients | Number of patients/cases | Leave blank if not reported |
| Batch size | Training batch size | Leave blank if not reported |
| Iterations | Training iterations or epochs | Use "X epochs" or "XK" for iterations |
| Architecture | Model architecture | e.g., ResNet-50, ViT-B, ViT-L |
| Parameters | Number of parameters | e.g., 86M, 632M |
| Embed dim | Output embedding dimension | e.g., 768, 1024, 2048 |
| Input size | Input image size at inference time (in pixels) | Usually 224 |
| Magnification | Leave blank | Will be added separately via add-magnification skill |
| Dataset | Training dataset names | e.g., TCGA, PAIP, proprietary |
| Links | GitHub and/or HuggingFace icons | See format below |
For slide-level models, the fields differ slightly (no Tiles column, has Patch size instead).
Use this format for links:
[<img src="https://raw.githubusercontent.com/FortAwesome/Font-Awesome/6.x/svgs/brands/github.svg" width="20">](GITHUB_URL)
[<img src="https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg" width="25">](HF_URL)
Mark vision-language models with (VL) after the name.
Models are ordered by release date. Find the correct position and insert the new row.
Cross-check extracted values:
For values that required inference or calculation (not directly stated in paper/code), add a brief explanation to NOTES.md. This provides transparency for how difficult-to-find values were derived.
Only document values that:
The order of models in NOTES.md should match their order in README.md (chronological by release date).
Keep notes concise—one bullet point per inferred value.
After adding the row, provide the user with a detailed summary of your findings. For every column in the row, include:
Format each field like this:
**Field Name**: [value]
- Quote: "[exact quote from paper/GitHub]" (Section X.X / GitHub README)
- Reasoning: [explanation of how you derived the value from the quote]
If a value was not found, explain:
If a value required inference or calculation (e.g., computing tiles from WSIs × tiles-per-WSI), show your work.
This transparency helps the user verify accuracy and catch any misinterpretations.
Note: Remind the user that magnification was left blank and can be added using the add-magnification skill.
| [ModelName](https://paper-url) | [Lab Name](https://lab-url) | :white_check_mark: | Jan 2024[\*](https://arxiv.org/abs/XXXX.XXXXXv1) | DINOv2 | 50K | 100M | 10K | 1024 | 100 epochs | ViT-B | 86M | 768 | 224 | | TCGA | [<img src="https://raw.githubusercontent.com/FortAwesome/Font-Awesome/6.x/svgs/brands/github.svg" width="20">](https://github.com/org/repo) |
K for thousands, M for millions, B for billionsINE suffix for "ImageNet epochs" if applicable** note in row if a value was inferred from other numbers