HYdsl/FinQA-Table-random-DeBERTa-Reranker

1---2license: apache-2.03language:4- en5base_model:6- naver/trecdl22-crossencoder-debertav37pipeline_tag: text-classification8tags:9- reranker10- cross-encoder11- financial-qa12library_name: transformers13---14 15# FinQA-Table-random-DeBERTa-Reranker16 17A passage reranker for the **HiREC** framework, fine-tuned from `naver/trecdl22-crossencoder-debertav3` on table data from the FinQA training set. General-purpose rerankers often fail to capture table-specific cues (titles, periods, indicators) that matter more than raw numerical values; this model is adapted to address that gap.18 19- 📄 Paper: [ACL 2025 Findings](https://aclanthology.org/2025.findings-acl.855/)20- 💻 Code: [LOFin-bench-HiREC](https://github.com/deep-over/LOFin-bench-HiREC)21 22## Training Data23 24Constructed from the **FinQA** training set, where each question is paired with an evidence page containing the gold table.25 26- **Positive passages:** tables located on the evidence page of each question.27- **Negative passages:** tables sampled from pages *other than* the evidence page within the same document (**random** negative sampling).28- For each positive, `n_neg = 8` negatives are drawn.29 30## Training Setup31 32- **Base model:** `naver/trecdl22-crossencoder-debertav3`33- **Objective:** Binary cross-entropy on `(query, passage)` pairs; the cross-encoder applies an internal sigmoid, producing relevance scores in [0, 1].34- Batch size: 128 / Epochs: 3 / Learning rate: 2e-735- **Hardware:** 1× NVIDIA GeForce RTX 409036 37## Citation38 39```bibtex40@inproceedings{choe-etal-2025-hierarchical,41    title     = {Hierarchical Retrieval with Evidence Curation for Open-Domain Financial Question Answering on Standardized Documents},42    author    = {Choe, Jaeyoung and Kim, Jihoon and Jung, Woohwan},43    booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},44    year      = {2025},45    url       = {https://aclanthology.org/2025.findings-acl.855/}46}47```