Skip to content
Language models

MedSAE for bge-large-en-v1.5

interpretable medical retrieval via sparse autoencoders

2 min readupdated

A collection of sparse autoencoders (SAEs — models that decompose dense vectors into a set of distinct, human-understandable features) from the doctolib-lab team, accompanying their ML4H 2026 paper. The model takes 1024-dimensional embeddings from the popular BAAI/bge-large-en-v1.5 encoder and maps them into a sparse code of 3072 latent features, with only k features active at a time (64 or 256). Unlike standard dense vectors, each active feature corresponds to a specific medical concept, and the k256_m3072 configuration even outperforms the base model in retrieval accuracy on the target domain. Useful for researchers and developers building search systems over English biomedical literature who need to see why a document matched a query.

Key features

  • Two ready-to-use checkpoints tailored for different goals: k64_m3072 for maximum feature interpretability and k256_m3072 for higher retrieval accuracy.
  • Retrieval runs via standard cosine similarity between sparse query and document codes, requiring no complex extra layers.
  • Trained exclusively on the CURE dataset (1,400 medical literature queries) using a combined MSE and InfoNCE loss function.
  • Weights are released in safetensors format alongside cfg.json under the non-commercial CC BY-NC 4.0 license.

What you can do

  • Build explainable search over English medical research papers by surfacing the query's active latent concepts alongside each retrieved document.
  • Use the k64_m3072 checkpoint in interpretability research to analyze which medical topics the BGE model encodes.
  • Drop in k256_m3072 sparse codes to replace standard BGE embeddings in biomedical and bioinformatics retrieval pipelines.

Example

import json, torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from sentence_transformers import SentenceTransformer

w = load_file(hf_hub_download("doctolib-lab/medsae-bge-large-en-v1.5", "k64_m3072/sae.safetensors"))
cfg = json.load(open(hf_hub_download("doctolib-lab/medsae-bge-large-en-v1.5", "k64_m3072/cfg.json")))
# После кодирования через BGE (с префиксом для query) применяем TopK + ReLU к (x - b_pre) @ encoder.T + bias
Read moreDetails
NDCG@10 (averaged across 3 seeds): on the CURE dataset, k256_m3072 scores 60.50% compared to 53.06% for the original BGE; on R2MED Medical Sciences, 29.05% vs. 27.78%; and on R2MED Bioinformatics, 28.10% vs. 27.04%.
On several out-of-domain datasets (NFCorpus, MIRIAD, R2MED Biology), the sparse codes trail the baseline BGE.
Limitations: English-only and restricted to scientific literature; the model has not been evaluated on clinical notes and is not intended for clinical decision-making.
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.0 GB0.6 GB
  • 4-bitweights 0.0 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # doctolib-lab/medsae-bge-large-en-v1.5 — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve doctolib-lab/medsae-bge-large-en-v1.5 --max-model-len 8192 --quantization fp8
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of bge-large-en-v1.5 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

doctolib-lab/medsae-bge-large-en-v1.5
Author
doctolib-lab
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments