bge-small-en-v1.5 fine-tuned on QASPER

BAAI/bge-small-en-v1.5 fine-tuned for two epochs with MultipleNegativesRankingLoss on 3,560 (question, gold evidence paragraph, BM25 hard negative) triples from the QASPER train split. On the QASPER test split (papers disjoint from training) it improves recall at a 1,024-token context budget by +0.149 [+0.127, +0.171] over the base model with paragraph-packed 256-token chunks. Query prefix: Represent this sentence for searching relevant passages: .

The app_bundle/ folder holds the precomputed test-split chunks and embeddings used by the demo app. Project, evaluation harness and full results: https://github.com/GFS418/rag-eval-harness

Downloads last month: 22

Safetensors

Model size

33.4M params

Tensor type

F32

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for GFS418/bge-small-qasper-ft

Base model

BAAI/bge-small-en-v1.5

Finetuned

(399)

this model