finalign / DATA.md
pdmdp's picture
Upload 22 files
c0f3416 verified
|
Raw History Blame Contribute Delete
2.09 kB

Data and models

The demo uses the FiQA retrieval collection distributed by BEIR, loaded through ir_datasets as beir/fiqa. Its dataset card specifies CC BY-SA 4.0. The original collection has 57,638 documents; filtering empty texts leaves 57,600.

The index bundle contains document texts, 250-word passage identifiers, BGE embeddings, a Terrier lexical index, and the development/test queries with their relevance judgments. These are derived artifacts of FiQA, distributed under CC BY-SA 4.0. The document IDs and text are retained; whitespace is collapsed only in passages. Each file has a SHA-256 digest in manifest.json.

The corpus is a historical benchmark containing public financial discussion, including posts unrelated to a visitor’s query. It is not a source of current market data or financial advice.

References:

  • Maia et al., WWW'18 Open Challenge: Financial Opinion Mining and Question Answering, 2018.
  • Thakur et al., BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models, arXiv:2104.08663, 2021.

The bundle’s models.json fixes the model revisions used for indexing and inference:

Role Model License
Passage and query encoder BAAI/bge-base-en-v1.5 MIT
Query expansion Qwen/Qwen2.5-1.5B-Instruct Apache 2.0
Reranking castorini/monot5-base-msmarco Not specified on the model card
monoT5 tokenizer google-t5/t5-base Apache 2.0

Weights are downloaded from their publishers at startup and are not included in the upload package. The monoT5 scoring implementation is provided by pyterrier_t5, the same library used by the notebook, pinned to a commit in requirements.txt.