| --- |
| license: cc-by-4.0 |
| tags: |
| - retrieval |
| - mathematics |
| - embeddings |
| pretty_name: RecipeMatching evaluation artifacts |
| --- |
| |
| # RecipeMatching evaluation artifacts |
|
|
| Per-query rank dumps and embedding caches behind the paper *Recipe-Matching, Not Equivalence* |
| (Ali Habibullah, Mohammad Alshiekh, Yazan Alshoibi, Salman Khan and Naeemullah Khan, 2026). |
| Code, data and results: <https://github.com/KAUST-Academy/recipe-matching-not-equivalence>. |
|
|
| These files are too large for the GitHub repository. Every number in the paper is already stored |
| there as a result JSON; the files here are what the analysis scripts need in order to recompute the |
| confidence intervals and the style analyses without re-running any model. |
|
|
| ## Contents |
|
|
| | Path | Files | Size | What it is | |
| |---|---|---|---| |
| | `results/ranks/*.ranks.jsonl` | 891 | 1.8 GB | One JSON line per query: its id and the rank of its gold document (fields vary by evaluation, e.g. `gold_rank`, `gold_sim`, `top10_ids`). Written by `scripts/eval_retrieve.py --dump-ranks`, `scripts/eval_crosslingual.py`, `scripts/eval_bm25.py` and `scripts/colbert_eval.py`. The repository's `results/ranks/*.summary.json` files are their aggregates. | |
| | `emb_cache/*.npz` | 16 | 4.4 GB | MathNet-Retrieve document (117,088) and query (15,000) embeddings of eight fine-tuned checkpoints, arrays `ids` and `emb`: `ctrl-cas-6145`, `ctrl-llm-6145`, `dose-paraphrase-6145`, `dose-style-6145`, `dose-unrelated-6145`, `fact-d4casnegs-6145-s42`, `fact-d4llmnegs-6145-s42` and `fact-d4llmnegsfull-6145-s42` (training records in `models/<name>/run_config.json` of the repository). | |
| | `upstream_versions/` | 2 | small | The Hugging Face revision of every base model and dataset the experiments downloaded, recorded on 2026-09-28, and the commit of the MIRB code. | |
|
|
| ## Use |
|
|
| Clone the GitHub repository, then place the files where the scripts expect them: |
|
|
| ```bash |
| hf download KAUSTAcademy/RecipeMatching_eval-artifacts --repo-type dataset --local-dir artifacts |
| cp artifacts/results/ranks/*.ranks.jsonl results/ranks/ |
| mkdir -p .emb_cache && cp artifacts/emb_cache/*.npz .emb_cache/ |
| ``` |
|
|
| The bootstrap, difference-in-differences and slice scripts (for example `scripts/bootstrap_stats.py`, |
| `scripts/dose_bootstrap.py`, `scripts/reference_control_did.py`, `scripts/xling_slice_analysis.py` and |
| `scripts/samelang_verdict.py`) read the rank dumps; `scripts/style_probe.py`, `scripts/style_within_class.py` |
| and `scripts/analyze_hits.py` read the embeddings. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{habibullah2026recipematchingequivalence, |
| title={Recipe-Matching, Not Equivalence}, |
| author={Ali Habibullah and Mohammad Alshiekh and Yazan Alshoibi and Salman Khan and Naeemullah Khan}, |
| year={2026}, |
| eprint={2609.31927}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.IR}, |
| url={https://arxiv.org/abs/2609.31927}, |
| } |
| ``` |
|
|
