Sgupta02/code-search-reranker-finetuned

Fine-tuned reranker (best checkpoint, step 300)

Fine-tuned version of Alibaba-NLP/gte-reranker-modernbert-base on hard-negative-mined query/document pairs from scikit-learn's API documentation.

Honest result: this does not beat the pretrained base model. Best checkpoint (step 300 of 1194, learning rate 2e-6) scored 0.534 nDCG@10 on a 26-query eval set, versus 0.556 for the untouched pretrained model. Included here as the actual artifact from a real fine-tuning attempt, not as a recommended model to use.

Full pipeline, evaluation methodology, and the stage-by-stage ablation (dense retrieval to hybrid search to reranking) that this was built on top of: https://github.com/sgupta2346/Code-Search-Finetuned-Reranker