RecipeMatching recipe arm (Qwen3-Embedding-0.6B)

A checkpoint from the paper Recipe-Matching, Not Equivalence (Ali Habibullah, Mohammad Alshiekh, Yazan Alshoibi, Salman Khan and Naeemullah Khan, 2026), shared so the paper's evaluations of this model can be re-run without retraining. It is not one of the three models the paper releases as its main artefacts. Code, data and results: https://github.com/KAUST-Academy/recipe-matching-not-equivalence.

What it is. The recipe arm of the paper's matched-budget experiment, also rung D1 of its rewriter-prompt ladder: Qwen3-Embedding-0.6B fine-tuned (all weights) on the 6,145 rows of data/llm_pairs/pairs.jsonl, LLM-written positives with, in nearly every row, near-miss negatives, produced with MathNet-Retrieve's published rewriter prompt, near-verbatim, by Qwen3-32B-AWQ and kept by a single LLM judge with no symbolic check. It shares every hyperparameter with the verified arm and differs from it only in the training file. Seed 42; the full training record, command line included, is run_config.json in this repository, and every training file named here is in the GitHub repository.

Results of this checkpoint (seed 42, R@1 / R@5 / R@10; the paper's main tables report eight-seed means (seeds 42–49), and its prompt-ladder table three-seed means):

Evaluation	R@1	R@5	R@10
MathNet-Retrieve easy tier (15,000 queries, 117,088 documents)	61.88	89.94	93.63
MathNet-Retrieve medium tier	9.06	55.18	66.79
MathNet-Retrieve hard tier	9.37	57.06	70.13
Cross-language duplicates (strict)	64.38	81.93	86.01

Usage. Encode queries with the model's query prompt, the setting every number above uses.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("KAUSTAcademy/RecipeMatching_RecipeArm_Qwen3-Embedding-0.6B")
q = model.encode(["Find all real x with x^2 = 2x."], prompt_name="query")
d = model.encode(["Determine every real solution of x^2 - 2x = 0."])
print(model.similarity(q, d))

Citation

@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}, 
}