EXCISE-checkpoints / README.md
mohammed933's picture
EXCISE release
c110648
|
Raw History Blame Contribute Delete
1.95 kB
metadata
license: mit
language:
  - en
library_name: peft
pipeline_tag: feature-extraction
tags:
  - information-retrieval
  - colbert
  - late-interaction
  - negation
  - exclusion
  - lora
  - peft
base_model:
  - lightonai/Reason-ModernColBERT
  - lightonai/GTE-ModernColBERT-v1
  - colbert-ir/colbertv2.0
pretty_name: EXCISE checkpoints

EXCISE checkpoints

Checkpoints of EXCISE: Query-Side Exclusion for Late-Interaction Retrieval, and of the systems the paper compares it with. Code and benchmark: github.com/DataScienceUIBK/EXCISE. Corpora: mohammed933/EXCISE-corpora.

Folder Content
excise/ per backbone, the detector (detector_<backbone>/deepfinder.pt, threshold in thr.txt) and the exclusion adapter (adapter_<backbone>/lora.pt); 25 MB
baselines/ the LoRA, full and query-only fine-tunes of each backbone
externals/ the fine-tuned external retrievers and rerankers

Pooled exclusion success@10 on X-BENCH (1,860 queries), with the frozen index for reference:

Backbone Detector Adapter EXCISE Frozen
Reason-ModernColBERT detector_reason adapter_reason 0.7731 0.2011
GTE-ModernColBERT-v1 detector_gte adapter_gte 0.7403 0.1608
ColBERTv2.0 detector_colbertv2 adapter_colbertv2 0.6995 0.2301

Each .pt stores the arguments it was trained with under args, and each fine-tuned baseline has a SELECTION.json with the learning rate chosen by five-fold cross-validation. The checkpoints run inside the repository's code:

git clone https://github.com/DataScienceUIBK/EXCISE && cd EXCISE
pip install -r requirements.txt
python scripts/download.py --what checkpoints               # excise/ only; --tier all for every folder
python -m excise.run --backbone reason --dataset excluir --system excise

MIT license.