facebookresearch/vizseq

An Analysis Toolkit for Natural Language Generation (Translation, Captioning, Summarization, etc.)

What it solves

VizSeq addresses the difficulty of analyzing text generation tasks across different modalities. It provides a unified way to visually inspect and evaluate predictions against references, supporting tasks like machine translation, summarization, image captioning, and speech translation.

How it works

It functions as a Python toolkit that takes multi-modal sources (text, image, audio, or video), text references, and text predictions as inputs. Users can perform their analysis through a Jupyter Notebook or a dedicated built-in Web App. Additionally, it includes a set of multi-process scorers to calculate common evaluation metrics efficiently.

Who it’s for

Researchers and developers working on text generation models who need to visually analyze their model outputs and calculate performance metrics across various input modalities.

Highlights

  • Broad Task Coverage: Supports text, image, audio, video, and multimodal inputs for tasks like OCR, speech recognition, and video description.
  • Extensive Metric Support: Includes n-gram-based metrics (BLEU, ROUGE, METEOR, etc.) and embedding-based metrics (BERTScore, LASER).
  • Flexible Deployment: Can be used as a library in Jupyter Notebooks or as a standalone web server.
  • High Performance: Uses multi-processing and multi-threading to accelerate the calculation of scorers.

Related

  • Project
    Vchitect/VBench

    VBench is an open‑source benchmark suite that automatically evaluates video‑generation models across dozens of quality dimensions (technical, aesthetic, and trustworthiness). It includes prompt collections, metric pipelines, human‑aligned scores, and public leaderboards for T2V, I2V, long‑video, and intrinsic‑faithfulness evaluation.

  • Project
    MinishLab/semhash

    SemHash is a lightweight Python library for fast, multimodal semantic deduplication, outlier filtering, and representative sample selection. It embeds data (text, images, etc.) using model2vec or custom encoders, leverages Vicinity’s ANN search for speed, and provides inspectable result objects. Ideal for cleaning large datasets, preventing train‑test leakage, and summarising data.

  • Project
    wavlab-speech/versa

    VERSA is an open‑source Python toolkit that bundles more than 90 speech‑ and audio‑quality metrics (perceptual, intelligibility, technical, and music‑specific). It supports plain files, Kaldi‑style lists, and distributed evaluation via Slurm, with optional GPU acceleration and visualisation. Install the core package, add optional dependency groups for heavy models, write a YAML config, and run the `versa/bin/scorer.py` CLI to obtain a multi‑metric evaluation report.

  • Project
    om-ai-lab/VLX-Seek

    VLX-Seek is a fine-grained perception VLM that replaces coordinate generation with region-token retrieval for more stable and efficient object localization in embodied AI.

  • Project
    alibaba/zvec

    Zvec is a lightweight, high-performance in-process vector database designed for embedding directly into applications to enable fast similarity and hybrid search without external server infrastructure.