impresso-project/multilingual-static-word-embeddings-demo

Multilingual Dictionary Explorer

This Space is a Gradio UI for the same lookup logic as query_multilingual_space.py: enter a source language and query word, then get translations to all other languages using FAISS, CSLS, and optional bidirectional consistency.

Required artifact files:

• —aligned_all.faiss
• —all_metadata.jsonl
• —config.json

The app does not use aligned_all.vec.

Runtime configuration

By default, the app scans the stage 6 prefix, downloads the newest artifact folder that contains config.json, and selects it in the UI:

s3://131-component-staging/multilingual-static-word-embeddings/stage-6/

Older artifact folders found under the same prefix are available from the Dictionary artifact dropdown. Switching artifacts reloads the aligned space and updates the displayed load time.

Set these Hugging Face Space secrets for S3-compatible storage:

• —SE_ACCESS_KEY
• —SE_SECRET_KEY
• —SE_HOST_URL

Optional environment overrides:

• —SPACE_ARTIFACT_S3_URI: exact artifact folder, for example s3://131-component-staging/multilingual-static-word-embeddings/stage-6/multilingual_dict_20260603_122323
• —SPACE_ARTIFACT_S3_PREFIX: prefix to scan for the newest multilingual_dict_* or multilingual_space_*.json
• —SPACE_DIR: local artifact folder, useful for local testing
• —ARTIFACT_CACHE_DIR: local cache directory, default /tmp/multilingual_space_artifacts

Defaults for top_k, min_score, csls_k, candidate_retrieval_k, csls_prefetch_k, and bidirectional consistency are read from config.json.