Modelpublic

jsanzolac/bpe_glove_300_lora_r300_qwen3_msmarco_ft

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes

Model Card

bpeglove300lorar300qwen3msmarco_ft

Stage 2 supervised fine-tuning of jsanzolac/bpe_glove_300_lora_r300_qwen3 on MS MARCO triplet-50 with InfoNCE-only (no MSE), 1 epoch, batch=256, K=6 mined hard negatives per anchor, τ=0.02.

Trainable: A.weight, B.weight only. E (frozen 300-d cl100k GloVe from jsanzolac/drifting-glove-distilled-r300) is excluded from the saved state dict.