This repository contains the full-parameter retriever checkpoint for XIR Competition 1170, based on Qwen3-Embedding-0.6B.
Method
The checkpoint is a full-parameter weight interpolation between the A10 baseline retriever and the g104 focus-training candidate:
checkpoint = 0.25 * A10 + 0.75 * g104
Only the official LRAT trajectories, the official BrowseComp-Plus offline corpus, and the original Qwen3-Embedding-0.6B weights were used. No external data, models, or generated external labels were used.
Local validation
A/B average MRR: 0.723194
A/B average Recall@1: 0.593750
Baseline A10 average MRR: 0.722926
The model is uploaded as full parameters in the repository root.
