This is the expand version of dleemiller/WordLlamaDetect model. Stacking two wordllama-based models to enhance the performace

• Support languages: 148

Training data (740k samples)
 │
 ▼
┌───────────────────────────────────┐
│ Phase 1: Base Models │
│ │
│ ┌─────────────┐ ┌─────────────┐ │
│ │LID Model 01 │ │ LID Model 02│ │
│ └──────┬──────┘ └──────┬──────┘ │
└─────────┼────────────────┼────────┘
 │ train each │
 │ independently │
 ▼ ▼
 lid_models[0] lid_models[1]
 │ │
 └───────┬────────┘
 │
 ▼
 collect_preds() → X: (N, 2*148) = (N, 296)

model1 logits model2 logits
 (N, 148) cat (N, 148)
 └────────┬──────────┘
 ▼
 (N, 296)
 │
 Linear(296 → 148) ← 296*148 = 43,808 params trained
 │
 ▼
 (N, 148) → CrossEntropy(y)

Evaluation results on Flores +

How to use code:

import sys
from pathlib import Path
from huggingface_hub import snapshot_download

# Download all files
local_dir = snapshot_download(repo_id="Bonkh/lid-stack-model-gemma_3_27b-gemma_3_300m")

# Load model.py
sys.path.insert(0, local_dir)
from model import LIDStack

# Load model
model = LIDStack.from_pretrained("Bonkh/lid-stack-model-gemma_3_27b-gemma_3_300m")

# Inference
print(model.predict("Hello, how are you?")) # → "eng_Latn"
print(model.predict(["Bonjour", "こんにちは"])) # → ["fra_Latn", "jpn_Jpan"]
print(model.predict("Xin chào", return_probs=True)) # → [("vie_Latn", 0.97)]

Base model