fenyo/vec2text-qwen3-fr-inverter-32

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

fenyo/vec2text-qwen3-fr-inverter-32 is a 0.8 billion parameter inverse text embedding model developed by Alexandre Fenyo. It reconstructs short French texts (4-32 tokens) from 1024-dimensional Qwen3-Embedding-0.6B vectors by projecting them into 16 pseudo-tokens and feeding them to a re-trained Qwen3-0.6B-Base decoder. This model serves as the initial inversion step in a vec2text chain, designed to assess data leakage risk from vector databases and test privacy protections.

Loading preview...

Overview

fenyo/vec2text-qwen3-fr-inverter-32 is a specialized inverse text embedding model designed to reconstruct short French texts (4-32 tokens) from their corresponding Qwen/Qwen3-Embedding-0.6B vectors. It acts as the crucial first step in a larger vec2text pipeline, aiming to reverse the embedding process. The model takes a 1024-dimensional, normalized embedding vector and projects it into 16 pseudo-tokens, which then guide a fully re-trained Qwen/Qwen3-0.6B-Base decoder to generate an initial text hypothesis.

Key Capabilities

  • Inverse Text Embedding: Reconstructs original French text from Qwen/Qwen3-Embedding-0.6B vectors.
  • Short Text Focus: Optimized for texts between 4 and 32 tokens (typically single sentences).
  • Pipeline Component: Designed to work in conjunction with the fenyo/vec2text-qwen3-fr-corrector-32 model for iterative refinement of the reconstructed text.
  • Privacy Assessment: Primarily used to measure data leakage risk from vector databases and test privacy protection mechanisms like de-identification or vector rotation.

Good for

  • Evaluating Vector Database Security: Assessing how much original text information can be recovered from embeddings.
  • Developing Privacy Protections: Testing the effectiveness of methods to safeguard sensitive information stored as embeddings.
  • Research in Inverse Text Modeling: Exploring the challenges and capabilities of reconstructing text from its latent representations.

Performance Highlights

When used as part of the full vec2text chain with a corrector, the model achieves significant reconstruction accuracy:

  • BLEU Score: Up to 72.4 (with 10 correction steps).
  • Exact Match: Up to 48% of texts reproduced identically.
  • Cosine Similarity: 0.985 between reconstructed and original text vectors.

Finetunes
1 models