STALI-paper2-adapters

STALI Paper 2 — DocHop-QA adapters

LoRA adapters and per-query DocHop-QA evaluations for A Late-Interaction Retriever Matches Graph-RAG on Multi-Hop Document Retrieval: A Matched-Reader Re-Evaluation.

Canonical repository: https://github.com/apertis-ai/multihop-doc-qa/tree/main/paper2

Contents

• adapters/a2_random_only/: primary A2 adapter, seed 42.
• adapters/a2_seed{1,2,3}/: additional A2 seeds.
• adapters/s1_stali/, s2_plain_lora/, s2_pure_hard_neg/, a2_small_control/: ablations.
• eval_results/: 500-query per-system evaluation outputs.
• eval/gold_n500.json: convenience copy of the GitHub gold file; SHA-256 7a35f83e0a38ac86da125db8ad3705295186619588ef9827752712d65ca5470d.

Load an adapter

from peft import PeftModel
from pylate.models import ColBERT

model = ColBERT(model_name_or_path="lightonai/GTE-ModernColBERT-v1")
model[0].auto_model = PeftModel.from_pretrained(
    model[0].auto_model,
    "theQuert/STALI-paper2-adapters",
    subfolder="adapters/a2_random_only",
)

The adapters target ModernBERT attention modules Wqkv and Wo with LoRA rank 16 and alpha 32. See the GitHub release for code, configurations, exact data provenance, and limitations.

License

Adapter weights and authored evaluation outputs are MIT licensed. The convenience gold file is derived from CC BY 4.0 DocHop-QA data and retains that upstream license and attribution.