Models¶

Pass the name in the first column to fairmedfm run --model. "Install" says what provides the model's dependencies: "runner" is the benchmark runner, [seg] its segmentation extra, and "separate" an additional install (see Installation).

Classification¶

All classification models support linear probing (--usage lp). Models marked CLIP-style also support clip-zs and clip-adapt.

`--model`	Model	CLIP-style	Install	Weights
`CLIP`	OpenAI CLIP	yes	runner	downloaded
`BLIP`	BLIP	yes	separate (LAVIS)	downloaded
`BLIP2`	BLIP-2	yes	separate (LAVIS)	downloaded
`BiomedCLIP`	BiomedCLIP	yes	runner	downloaded
`PubMedCLIP`	PubMedCLIP	yes	runner	downloaded
`MedCLIP`	MedCLIP	yes	runner	downloaded
`PLIP`	PLIP (pathology)	yes	runner	downloaded
`SigLIP`	SigLIP	yes	runner	downloaded
`SigLIP2`	SigLIP 2	yes	runner	downloaded
`MedSigLIP`	MedSigLIP	yes	runner	gated
`CONCH`	CONCH (pathology)	yes	separate	gated
`DINOv2`	DINOv2		runner	downloaded
`DINOv3`	DINOv3		runner	gated
`AIMv2`	AIMv2		runner	downloaded
`RADDINO`	RAD-DINO (chest X-ray)		runner	downloaded
`MedGemma`	MedGemma vision tower		runner	gated
`UNI2`	UNI2-h (pathology)		runner	gated
`Virchow2`	Virchow2 (pathology)		runner	gated
`ProvGigaPath`	Prov-GigaPath (pathology)		runner	gated
`RETFound`	RETFound (ophthalmology)		runner	gated, local file
`Merlin`	Merlin (3D CT)		separate	downloaded
`MedLVM`	LVM-Med		runner	`pretrained/`
`C2L`	C2L		runner	`pretrained/`
`MedMAE`	MedMAE		runner	`pretrained/`
`MoCoCXR`	MoCo-CXR		runner	`pretrained/`

downloaded: fetched from Hugging Face or the model's source on first use.
gated: request access on the model's Hugging Face page, then huggingface-cli login or set HF_TOKEN.
pretrained/: from pretrained.zip (see Run experiments).
RETFound ships .pth files: download it after access is granted and set pretrained_path in configs/models/RETFound.json in your working directory.

Segmentation¶

All segmentation models use --task seg --usage seg2d with a --prompt.

`--model`	Model	Install	Checkpoint
`SAM`	Segment Anything (ViT-B)	`[seg]`	`--sam_ckpt_path`
`MedSAM`	MedSAM	`[seg]`	`--sam_ckpt_path`
`MobileSAM`	MobileSAM	`[seg]`	`--sam_ckpt_path`
`TinySAM`	TinySAM	`[seg]`	`--sam_ckpt_path`
`SAMMed2D`	SAM-Med2D	`[seg]`	`--sam_ckpt_path`
`FT-SAM`	FT-SAM	`[seg]`	`--sam_ckpt_path`
`SAM2`	SAM 2	`[seg]` + SAM2	`--sam_ckpt_path`, `--sam2_model_cfg`
`MedSAM2`	MedSAM2	`[seg]` + SAM2	`--sam_ckpt_path`, `--sam2_model_cfg`
`SAM3`	SAM 3	`[seg]` + SAM3	downloaded (gated)
`MedicalSAM3`	Medical SAM3 (2D, box prompts only)	`[seg]` + SAM3	`--sam_ckpt_path` to `checkpoint_2D.pt`

SAM3 requires Python 3.12+ and PyTorch 2.7+; use a separate environment if needed. MedicalSAM3 keeps detections above confidence 0.1 and uses the highest-scoring mask; an empty detection produces an empty mask.

Your own model¶

To evaluate a model that is not listed, you do not need to add it to FairMedFM: run it yourself and score its predictions with fairmedfm score.