Download README.md from pdmdp/finalign: direct link, hf CLI and curl.
- Browser
- Download file 1.65 kB
-
https://huggingface.co/spaces/pdmdp/finalign/resolve/main/README.md
- Command line
-
hf download hf://spaces/pdmdp/finalign/README.md
-
curl -L -o README.md https://huggingface.co/spaces/pdmdp/finalign/resolve/main/README.md
title: FinAlign
emoji: 🔎
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 6.32.0
python_version: 3.12.12
app_file: app.py
pinned: true
license: mit
short_description: Compare lexical and neural search on financial questions.
models:
- BAAI/bge-base-en-v1.5
- Qwen/Qwen2.5-1.5B-Instruct
- castorini/monot5-base-msmarco
datasets:
- BeIR/fiqa
tags:
- information-retrieval
- finance
- query-expansion
- reranking
startup_duration_timeout: 1h
hf_oauth: true
FinAlign
Compare BM25 with BGE passage retrieval on the FiQA corpus. Turn on Qwen query expansion, lexical fusion and monoT5 reranking to inspect how the ranking and latency change.
The demo follows the FinAlign course project. It uses the complete nonempty corpus: 57,600 documents split into 66,113 passages. Query variants are intermediate retrieval inputs, not answers to the visitor.
Rankings and timings are computed on demand. Quality metrics appear only when a question exactly matches a uniquely identified FiQA query with relevance judgments. The historical notebook results are separate from these runs.
See DEPLOY.md for setup, EXPERIMENTS.md for ablations and DATA.md for data and model provenance.
Original project: Matteo Pietro Di Pilato, Lorenzo Cinquemani and Lorenzo Marchioni. Matteo implemented the retrieval pipelines and ran the original evaluation. The interactive version is a later extension.
Code is MIT licensed. The bundled FiQA collection and derived index are distributed under CC BY-SA 4.0; pretrained models retain their respective licenses.