LEMMO: Large ElectroMagnetic MOdel

This repository provides the implementation of LEMMO. It converts raw complex I/Q samples into continuous signal tokens and conditions a language model to answer electromagnetic-signal questions.

Two inference variants are implemented:

Variant	Intended use	IQ length	Signal-language interface
LEMMO_Recognition	Four-task open-ended recognition	128 to 5,000,000 (validated lengths below)	16 question-conditioned signal queries
LEMMO_Description	Long-form electromagnetic signal description	16 to 8,000,000	continuous 1M-point chunk encoding and 64 question-conditioned signal queries

Repository layout

.
├── configs/
├── lemmo/
├── infer.py
└── requirements.txt

Requirements

The reference runtime was validated with Python 3.13, CUDA on NVIDIA H100, PyTorch 2.9.1, Transformers 5.13.1, and Flash Linear Attention 0.5.1. Other CUDA accelerators and dependency versions have not yet been validated.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

If TileLang needs an explicit CUDA toolchain, set it before inference:

export LEMMO_TOOLCHAIN=/path/to/cuda-toolchain

Weights

LEMMO_Recognition: https://drive.google.com/file/d/1MV4rREcGGVfflDzI6g5qpnfjMsoQIMd1/view?usp=drive_link LEMMO_Description: https://drive.google.com/file/d/13P4pv_loh59yWEeF8Y29EZHiPk2OyZG_/view?usp=drive_link

After downloading the weights, create a local weights/ directory and place them at:

weights/LEMMO_Recognition.pt
weights/LEMMO_Description.pt

Input format

Both variants accept:

• a NumPy .npy signal with real shape [L, 2] or complex shape [L], or a two-column .csv file;
• a sampling rate in Hz; LEMMO_Description requires a finite positive value;
• acquisition-setting text for LEMMO_Description;
• a natural-language question.

LEMMO_Recognition additionally requires a task adapter name:

• amc: modulation recognition;
• interference: interference recognition;
• uav: UAV-device recognition;
• wtc: wireless-technology recognition.

LEMMO_Recognition has been validated at lengths 128, 256, 976, 1024, 4096, 16368, and 5000000. LEMMO_Description accepts any length from 16 through 8000000.

Quick start

LEMMO_Recognition:

python infer.py \
  --model LEMMO_Recognition \
  --config configs/LEMMO_Recognition.example.json \
  --signal /path/to/iq.npy \
  --sample-rate 20000000 \
  --task wtc \
  --question "这段信号采用了哪种无线通信技术？"

LEMMO_Description:

python infer.py \
  --model LEMMO_Description \
  --config configs/LEMMO_Description.example.json \
  --signal /path/to/iq.npy \
  --sample-rate 20000000 \
  --input-setting "信号采样率为20 MHz，观测时长由IQ点数计算。" \
  --question "请描述这段信号的关键时域、频域和调制特征。"