presto-arabic-search-models

PrestoAI Arabic E-Commerce Search - system artifacts

Model and reproduction artifacts for team Looooop Gain03, final public leaderboard 0.69413 nDCG@10 on the PrestoAI Arabic E-Commerce Search Retrieval Challenge, across eighteen submissions each predicted in advance to within +-0.0014 by a local harness.

Code: https://github.com/TahaElkhazmi/presto-arabic-search Setup and step-by-step run instructions: README.md in that repository (four paths: verify the score with no GPU, run inference, rebuild the submission from these weights, or retrain from scratch). Technical report: TECHNICAL_REPORT.md, which carries the full experimental record including nineteen diagnosed negative results.

When downloading, keep these local paths - the scripts expect them:

Contents

Pipeline

BM25 (normalized Arabic, ascending-id tie-break) and the LoRA dense retriever each supply top-100 candidates; these are fused with weighted reciprocal-rank fusion (k=60, lexical:dense 3:1) into a pool and a prior. The cross-encoder scores every pool pair and its sigmoid outputs are temperature-recalibrated at T=3.2. The final score is

0.50*CE + 0.50*prior + 0.35*match-position + 0.05*first-token-position
        + 0.05*highest-IDF-token-position + 0.05*prefix/typo-position
        + 0.04*proximity-bucket

A confidence-adaptive near-tie band (eps 0.030, widening to 0.050 when the CE top1-top3 margin < 0.10, and 0.080 over ranks 1-5) then re-sorts low-information regions by a five-level lexicographic key: char-3gram query-in-name containment computed after stripping eleven Libyan-dialect filler tokens, then match position, then first-token position, then last-token position, then ascending product id.

Reranker training stages

1. Arabic IR warm-up - BCE on 40k Arabic MS MARCO pairs.
2. Synthetic corpus-grounded curriculum - LambdaLoss on 60k groups generated from the competition catalog itself (query = size/unit-stripped name prefix; positive = the source product; hard negatives = char-3gram neighbour products the query lexically discriminates against). Labels are grounded by construction, addressing a measured 8-17% grounding rate in the provided labels.
3. Provided-pairs listwise fine-tune - LambdaLoss over per-query groups.

Pure-ordering quality by stage (identical pools/slice): pointwise BCE 0.524 < zero-shot 0.561 < listwise 0.593 < +warm-up 0.603 < +synthetic curriculum 0.621. A zero-shot 4B model of the same family scores 0.554.

Provenance note

synth_stage15_adapter (the Stage-2 LoRA) was deleted during cleanup after its merge completed. The merged Stage-2 checkpoint is the ground-truth artifact and the training script regenerates the stage from scratch; the adapter-by-adapter chain from the stock base is therefore not replayable for that stage. All other merges are verified at the state-dict level in fp32 against base + delta.

Licensing

Base checkpoints (Qwen3-Embedding-0.6B, Qwen3-Reranker-0.6B) are Apache-2.0; these derivatives inherit that license. Training used the competition-provided data, the public Arabic MS MARCO corpus, and the competition catalog. No benchmark relevance judgment contributed a gradient to any weight in these artifacts.