Haiku — Trimodal (CODEX + H&E + Text) Retrieval Model

This repo bundles a fine-tuned Haiku checkpoint together with the tokenizer and marker assets needed to run inference without any additional downloads from xiangjx/musk or microsoft/BiomedNLP-BiomedBERT-*.

Contents

• haiku_state_dict.pt — model weights (CODEX + H&E + Text encoders + projections)
• config.json — architecture config + marker lists
• tokenizer/ — BiomedBERT tokenizer files (+ bert config)
• esm_embeddings/ — per-biomarker ESM embeddings (also embedded in state_dict; kept here for downstream use)
• vocab.pkl — marker vocabulary

Quick start

from models import Haiku

model, tokenizer, marker_embedding = Haiku.from_pretrained(
    "zhihuanglab/Haiku",
    device="cuda",
    token="hf_...",  # omit if HF_TOKEN / hf auth login is set
)
model.eval()