AT A GLANCE
Evaluation dashboard
30 monolingual query-to-passage tests
LEADERBOARD
Model comparison
EXPERIMENT INSIGHT
What the scores reveal
PERFORMANCE BY SEGMENT
Where retrieval succeeds or fails
Small subgroup sample sizes are displayed with each result. Comparisons across tiny groups can be unstable.
EVALUATION REGISTRY
Neural candidates for the next phase
These models are configured in the reproducible Python script but have not been executed in this benchmark. They are never included in the ranking.
EXPLAINABILITY LAB
Inspect every retrieval decision
Every case contains a question and three candidate passages.
EVALUATION CASES
Choose an experiment
SCORE EXPLANATION
Passage ranking
REPRODUCIBLE RESEARCH
Methodology and limitations
Transparent by design, with the data and runner included.
01 / DATASET
Research Retrieval Eval Mini
48 manually curated query-to-passage tasks: 30 within-language examples and 18 cross-language examples. Every query has one positive passage, one difficult negative passage and one easier negative passage.
English, Spanish and Portuguese; five domains: research, artificial intelligence, business, education and Latin America.
02 / METHODS
Four measured baselines
Character TF-IDF (3 to 5 character n-grams), word TF-IDF (1 to 2 word n-grams), hybrid TF-IDF (60% character and 40% word), and BM25.
Embeddings are future candidates, not secretly inferred scores.
03 / METRICS
How to interpret the numbers
Accuracy@1: the fraction of queries where the positive passage ranks first. MRR: the average inverse rank of the positive passage.
Scores are based on three candidates per question. This is a controlled diagnostic, not open-web search evaluation.
04 / LIMITATIONS
What we cannot conclude
The curated dataset is small. Content and negative examples may contain lexical patterns that favor certain methods. Some cross-language cognates make character matching unusually effective.
This does not establish broad model superiority, statistical significance, generalization or reliable performance on unseen production data.
RUN IT YOURSELF
Reproducibility in four commands
These files are included in your published Hugging Face repository. No Python is required to view the dashboard.
pip install -r requirements-benchmark.txtpython benchmark.py --models lexical --split monolingualpython benchmark.py --models lexical --split cross-languagepython benchmark.py --models all --split cross-languageThe last command attempts neural models and needs additional optional dependencies and downloaded weights. Output is written to your local JSON files, not automatically uploaded to the Space.