Modelpublic

xling-latent-intervention/latent-intervention-qwen3-8b

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes

Model Card

# Cross-Lingual Latent-Space Intervention — Qwen3-8B (layer 28)

Trained language-specific decoder from the cross-lingual autoencoder used in the paper "Latent-Space Intervention for Cross-Lingual Factual Consistency: Consistency Improvements without Accuracy Drops" (Findings of EMNLP 2026).

This is the final (phase-2) model: a shared encoder with language-specific decoders, trained on parallel hidden states extracted from layer 28 of Qwen/Qwen3-8B.


Base model	Qwen/Qwen3-8B
Layer	28
Latent dim	256
Pooling	mean
Loss	Huber
Learning rate	1e-4
Languages	Arabic, English, Dutch, Russian, Chinese
Training data	Aligned TED Talk transcripts

## Files

—best_model.pth — encoder + language-specific decoders (PyTorch state dict).

## Usage

bash

  huggingface-cli download xling-latent-intervention/latent-intervention-qwen3-8B best_model.pth --local-dir ./models/Qwen3-8B                       
                                                                                                                                                                 
  Pass it to the experiment scripts via --ae_model_path (--ae_path for evaluation).                                                                              
  See the code repository                                                                                                                                        
  for the autoencoder_nn.py architecture and loading code.

## Citation

bibtex

  @inproceedings{ghorbanpour2026latent,                                                                                                                          
      title     = "Latent-Space Intervention for Cross-Lingual Factual Consistency: Consistency Improvements without Accuracy Drops",                            
      author    = "Ghorbanpour, Faeze and Fierro, Constanza and Fraser, Alexander and S{\o}gaard, Anders",                                                       
      booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2026",                                                                       
      year      = "2026",                                                                                                                                        
      publisher = "Association for Computational Linguistics",                                                                                                   
  }