slideflow-labs/mettle-1.5

METTLE 1.5

METTLE 1.5 is Slideflow Labs' robustness fine-tune of H-optimus-0 for histopathology feature extraction. It is designed to improve representation consistency across slide scanners and staining variation while retaining biologically useful morphology.

The earlier checkpoint is METTLE 1 (pre-release).

Property	Value
Architecture	ViT-g/14 with four register tokens and a CLS refinement head
Input	224 × 224 RGB tile
Default output	1,536-dimensional refined CLS embedding
Optional output	3,072-dimensional refined CLS + mean spatial-patch embedding
Weights	model.safetensors; original tensor values and precision preserved

The spatial-patch mean excludes the CLS and register tokens. The head dimensions are specified in config.json. Source and packaged weight checksums are recorded in weights_manifest.json.

Usage

Dependencies are listed in requirements.txt.

import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

model_id = "slideflow-labs/mettle-1.5"
device = "cuda" if torch.cuda.is_available() else "cpu"
processor = AutoImageProcessor.from_pretrained(model_id, use_fast=False)
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
).to(device).eval()

image = Image.open("tile.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt").to(device)

with torch.inference_mode():
    # Use "cls" for 1,536 features, or "cls_mean" for 3,072 features.
    embeddings = model(**inputs, feature_view="cls_mean").pooler_output

The processor resizes directly to 224 × 224 with bicubic interpolation, rescales RGB intensities by 1/255, and normalizes with mean [0.707223, 0.578729, 0.703617] and standard deviation [0.211883, 0.230117, 0.177517].

Packaging verification

All 577 saved tensors match the source checkpoint byte for byte. The saved package loads through AutoModel without missing or unexpected weights. On a synthetic RGB tile, both feature views match a reference implementation using the original checkpoint to a maximum absolute difference of 1.79e-7 in CPU float32 inference. See verification.json for the artifact identities and test results. The tested runtime used PyTorch 2.13.0, timm 1.0.29, and Transformers 4.57.1.

Compatibility

The Hugging Face loader uses backbone.* tensor names. Consumers that directly load the original flat .pt checkpoint require their existing checkpoint format or an explicitly compatible safetensors loader.

Scope

This is a tile feature extractor, not a standalone diagnostic system. This model card makes no benchmark or clinical-performance claims for this checkpoint.