RecipeMatching V2 (Qwen3-Embedding-0.6B)

One of the three models released with 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.

What it is. The paper's mixed-supervision model V2: Qwen3-Embedding-0.6B fine-tuned on a mixture of 30,323 rows from four sources (8,747 computer-algebra-verified pair rows with negatives capped at two per row, 12,000 problem-to-solution rows as replay, 4,576 cross-transform positives and 5,000 same-topic mid-difficulty negatives; data/phase2/mixed_pairs.jsonl in the repository), with the task instruction below attached to queries during training. Seed 42; the full training record is run_config.json in this repository, and the training command is in the GitHub README.

Results of this checkpoint (seed 42, R@1 / R@5 / R@10, queries encoded with the instruction):

Evaluation	R@1	R@5	R@10
MathNet-Retrieve easy tier (15,000 queries, 117,088 documents)	21.70	94.48	97.81
MathNet-Retrieve medium tier	2.71	67.25	82.06
MathNet-Retrieve hard tier	0.01	7.02	32.03
Cross-language duplicates (strict)	74.30	89.31	93.13

Usage. Encode queries with the task instruction, not the model's stored query prompt (which is the base model's default and not what V2 was trained with).

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("KAUSTAcademy/RecipeMatching_V2_Qwen3-Embedding-0.6B")
instruction = "Instruct: Given a math problem, retrieve problems that are mathematically equivalent to it.\nQuery:"
q = model.encode(["Find all real x with x^2 = 2x."], prompt=instruction)
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}, 
}