By 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 Model to README
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.
Prerequisites
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.
Workflow
Step 1: Determine the Table
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 |
Step 2: Extract All Fields
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).
Step 3: Format the Row
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.
Step 4: Insert in Chronological Order
Models are ordered by release date. Find the correct position and insert the new row.
Step 5: Verify Accuracy
Cross-check extracted values:
- Re-read relevant paper sections to confirm numbers
- Verify GitHub README matches paper claims
- Ensure batch size/epochs refer to main model training, not ablations
- Check that patient counts are for pretraining data, not evaluation sets
Step 6: Document Inferred Values in NOTES.md
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:
- Were inferred from multiple sources
- Required combining information from different parts of the paper
- Involved checking upstream repositories or external sources
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.
Step 7: Report Findings to User
After adding the row, provide the user with a detailed summary of your findings. For every column in the row, include:
- Value: The value you added (or state that you couldn't find it)
- Source quote: A direct quote from the PDF or GitHub that supports this value
- Reasoning: Brief explanation of how you determined this value from the quote
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:
- What sections you checked
- Why the information appears to be unreported
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.
Example Row
| [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) |
Conventions
- Round WSIs, tiles, patients to 2 significant figures (e.g., 14,325,848 → 14M)
- Use
Kfor thousands,Mfor millions,Bfor billions - Bold model name and WSI count if trained on >100K slides
- Use
INEsuffix for "ImageNet epochs" if applicable - Leave cells blank (not "N/A") for unreported values
- Add
**note in row if a value was inferred from other numbers