Jiarui-Wang

MemSFT-Qwen3-Bio-Memory-4B

README

License: apache-2.0

Introduction

MemSFT specializes modern large language models with an external parametric memory. This checkpoint contains the Qwen3-4B memory trained on Biology-Instructions. The memory learns to approximate retrieval-based teacher distributions over domain SFT data. At each decoding step, a learned token-level router combines the next-token distributions of the frozen base model and memory. This checkpoint is an auxiliary memory, not a standalone chat model. Its key advantages are:

Plug-and-Play: Attaches to a frozen backbone without modifying its parameters or architecture.
Strong Specialization: Improves domain performance with negligible degradation in general capabilities.
Cross-Scale Reuse: The MemSFT design supports reuse across compatible Qwen3 backbones without retraining the memory for each backbone.

Quick Start

This example pairs the 4B memory with the Qwen3-14B backbone used in the paper. It requires a CUDA GPU with sufficient memory to load both models in BF16.

1. Install

bash
git clone https://github.com/LUMIA-Group/MemSFT.git
cd MemSFT

conda create -n memsft-generate python=3.10 pip -y
conda activate memsft-generate
python -m pip install -e .
python -m pip install \
  "torch>=2.4,<2.7" \
  "transformers==4.51.3" \
  "huggingface-hub==0.35.3" \
  "accelerate>=0.34,<2"

2. Load the base, memory, and router

python
from pathlib import Path

import torch
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer

from memsft.router.adaptive_memdec import AdaptiveMemoryDecoder

device = torch.device("cuda:0")
base_id = "Qwen/Qwen3-14B"
memory_id = "Jiarui-Wang/MemSFT-Qwen3-Bio-Memory-4B"
router_repo = "Jiarui-Wang/MemSFT-Qwen3-Routers"
router_subdir = "Qwen3-14B-Bio-M4B-Router"

router_root = snapshot_download(
    repo_id=router_repo,
    revision="v1.0.0",
    allow_patterns=[f"{router_subdir}/*"],
)
router_path = str(Path(router_root) / router_subdir)

tokenizer = AutoTokenizer.from_pretrained(
    base_id,
    revision="40c069824f4251a91eefaf281ebe4c544efd3e18",
)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision="40c069824f4251a91eefaf281ebe4c544efd3e18",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).to(device).eval()
memory = AutoModelForCausalLM.from_pretrained(
    memory_id,
    revision="v1.0.0",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).to(device).eval()

vocab_size = len(tokenizer)
base.resize_token_embeddings(vocab_size)
memory.resize_token_embeddings(vocab_size)
base.requires_grad_(False)
memory.requires_grad_(False)
model = AdaptiveMemoryDecoder(
    base_lm=base,
    knn_generator=memory,
    router_path=router_path,
    router_device=device,
).eval()
model.set_tokenizer(tokenizer)

3. Generate

python
sequence = (
    "MKSILIEKPNQLAIVEREIPTPSAGEVRVKVKLAGICGSDSHIYRGHNPFAKYPRVIGHEFFGVIDAV"
    "GEGVESARVGERVAVDPVVSCGHCYPCSIGKPNVCTTLAVLGVHADGGFSEYAVVPAKNAWKIPEAVA"
    "DQYAVMIEPFTIAANVTGHGQPTENDTVLVYGAGPIGLTIVQVLKGVYNVKNVIVADRIDERLEKAKE"
    "SGADWAINNSQTPLGEIFTEKGIKPTLIIDAACHPSILKEAVTLASPAARIVLMGFSSEPSEVIQQGI"
    "TGKELSIFSSRLNANKFPIVIDWLSKGLIKPEKLITHTFDFQHVADAISLFEQDQKHCCKVLLTFSE"
)
prompt = (
    r"<PROTEIN> "
    + sequence
    + r" </PROTEIN> What is the EC number associated with the enzymatic "
    r"function of this protein? Please put the final enzyme within \boxed{} "
    r"using an EC number such as ECx.x.x.x, and separate multiple entries "
    r"with commas."
)

messages = [{"role": "user", "content": prompt}]
prompt_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt_text, return_tensors="pt").to(device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        do_sample=False,
        max_new_tokens=32,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )

answer = tokenizer.decode(
    output_ids[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(answer)

Example output:

text
\boxed{EC1.1.1.-}

UniProtKB annotates Escherichia coli K-12 RspB (P38105) with EC 1.1.1.-. For comparison, using the same prompt and deterministic generation configuration, Qwen3-14B alone predicts EC 4.2.1.22. The outputs were reproduced in BF16 on NVIDIA A800 80GB GPUs.

Performance

The memory-scale experiment uses the same Biology-Instructions training data, Qwen3-8B retrieval teacher, and frozen Qwen3-14B backbone across memory sizes.

Table with columns: Configuration, Memory size, Biology-Instructions ↑
Configuration	Memory size	Biology-Instructions ↑
Qwen3-14B	—	6.64
Qwen3-14B + MemSFT	4B	37.12

The paper reports a General average range of 83.23–83.62 across the 1.7B, 4B, and 8B memory configurations, rather than a separate General average for this checkpoint.

Compatible Pairing

base: Qwen/Qwen3-14B
memory: Jiarui-Wang/MemSFT-Qwen3-Bio-Memory-4B
router: Jiarui-Wang/MemSFT-Qwen3-Routers/Qwen3-14B-Bio-M4B-Router

MemSFT prefers tensor-only .safetensors router checkpoints. Legacy .pt checkpoints should be loaded only from trusted sources; the MemSFT loader uses PyTorch's restricted weights_only=True mode for compatibility.

Intended Use and Limitations

This checkpoint is intended for reproducing the MemSFT memory-scale experiment and for augmenting Qwen3-14B on Biology-Instructions tasks. It should be used with the matching Qwen3-14B/Bio/M4B router. This 4B memory was not evaluated with the other backbone sizes in the paper. Performance outside the evaluated pairing and domain has not been established.

License

This MemSFT checkpoint is released under the Apache License 2.0. Upstream models, software, and datasets remain subject to their respective licenses and terms.

Citation

If you find MemSFT helpful in your research, please consider citing:

bibtex
@misc{wang2026memsftmitigatingalignmenttax,
      title={MemSFT: Mitigating Alignment Tax with an External Parametric Memory},
      author={Jiarui Wang and Xiang Shi and Jiaqi Cao and Rubin Wei and Xiquan Wang and Hao Sun and Jingzhi Wang and Zhiqi Yang and Qipeng Guo and Bowen Zhou and Zhouhan Lin},
      year={2026},
      eprint={2607.25614},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.25614}, 
}

Contact

For questions and discussions, feel free to email wangjiarui1@sjtu.edu.cn.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Jiarui-Wang

Model Tree

Base

Qwen/Qwen3-4B

Fine-tuned

this model

Input Modalities

Text

Output Modalities