Model Card — BGE-Reranker-VietFinance

Overview

This is a cross-encoder reranker finetuned from BAAI/bge-reranker-v2-m3 to score (query, passage) pairs for retrieval reranking in Vietnamese financial/news search systems.

Intended Use

• Primary: improve Hit@K by re-ranking candidate passages produced by an upstream retriever (BM25 + embedding-based).
• Not for: standalone generation, non-Vietnamese domains, or high-stakes automated decisions without human review.

Essential Statistics

• Base model: BAAI/bge-reranker-v2-m3
• Embedding model (used for retrieval/hard-negative mining): BAAI/bge-m3
• Max sequence length for reranking: 1536 tokens (inputs longer than this are truncated)
• Retrieval strategy: temporal-aware hybrid (BM25 + dense embeddings with temporal boosting)
• Saved artifacts in this folder: model.safetensors, tokenizer.json, tokenizer_config.json, config.json.

Evaluation (concise)

• Procedure: retrieve candidate passages (temporal-aware hybrid) → rerank with cross-encoder → compute Hit@K for K ∈ {1,3,5,10,20}.
• Numeric results are saved in run outputs (summary CSVs / JSONL); include them here if you want the actual Hit@K values embedded.

Limitations & Risks

• Domain-specific: optimized for Vietnamese financial/news passages; generalization outside this domain/language is uncertain.
• Retrieval dependency: reranker cannot recover gold passages not present among retrieval candidates.
• Truncation risk: 1536-token truncation may drop important context for long passages.
• Data & license: dataset provenance and license are not specified here — verify before public distribution.

Bias & Safety

• Model reflects biases in the source news corpus (topic/regional biases).
• Temporal heuristics can misinterpret ambiguous locale-specific dates and cause incorrect boosts.
• Do not rely on reranker outputs alone for automated financial, legal, or medical decisions.

Quick usage (inference)

Load this checkpoint with AutoTokenizer / AutoModelForSequenceClassification, tokenize (query, passage) pairs, score in eval mode, and sort candidates by descending score (higher = more relevant).

License & Citation

• License: not specified in the checkpoint — confirm before redistribution.
• Cite the base models BAAI/bge-reranker-v2-m3 and BAAI/bge-m3 when reporting results.