README.md

Baseline Guide

This document contains baseline commands and related notes.

Dense Retriever

python dense_retriever.py --dataset_name covid

Cross Encoder

python cross_encoder.py \
  --dataset_name covid \
  --llm_budget 50

Pointwise LLM

pointwise_llm.py requires a prebuilt relevance score cache.

python pointwise_llm.py \
  --dataset_name covid \
  --llm_name unsloth/Qwen3-14B-unsloth-bnb-4bit \
  --llm_budget 50

Build a Relevance Score Cache

BAGEL stores LLM relevance scores in a reusable cache.
This cache is optional for BAGEL, but required for pointwise reranking.

python pointwise_label.py \
  --dataset_name covid \
  --llm_name unsloth/Qwen3-14B-unsloth-bnb-4bit \
  --top_k 50

This scores the top-k passages returned by the dense retriever.

Listwise LLM (RankGPT)

python listwise_llm.py \
  --dataset_name covid \
  --llm_name unsloth/Qwen3-14B-unsloth-bnb-4bit \
  --llm_budget 50 \
  --window_size 50