bactrianus/HotpotQA-Paragraph-Retriever-Llama-3-8B-Instruct
The bactrianus/HotpotQA-Paragraph-Retriever-Llama-3-8B-Instruct model, developed by bactrianus, is a Llama 3 8B Instruct-based causal language model specifically fine-tuned for relevant-paragraph selection in HotpotQA-style research. It excels at identifying paragraph titles containing evidence for answers from a fixed set of candidates, serving as an upstream component in modular question-answering pipelines. This model is designed for fixed-candidate paragraph relevance selection and analysis of recall and error propagation.
Loading preview...
Model Overview
The bactrianus/HotpotQA-Paragraph-Retriever-Llama-3-8B-Instruct is a specialized Llama 3 8B Instruct model, fine-tuned by bactrianus, designed for paragraph selection within the HotpotQA framework. Unlike traditional retrievers, this model does not query an external corpus but rather selects relevant paragraphs from a pre-supplied fixed set of candidates. Its primary function is to receive a question and candidate paragraphs, then generate the titles of paragraphs predicted to contain the necessary evidence for an answer, acting as an intermediate interface for subsequent sentence selectors or readers.
Key Capabilities
- Fixed-Candidate Paragraph Selection: Identifies relevant paragraphs from a given set for HotpotQA-style research.
- Modular Pipeline Component: Functions as an upstream element in a selector-reader pipeline for complex question answering.
- Analysis Tool: Useful for analyzing paragraph recall and understanding error propagation from paragraph selection to sentence-level tasks.
Intended Use Cases
- HotpotQA Research: Specifically designed for research involving fixed-candidate paragraph relevance selection in HotpotQA-style datasets.
- Modular QA Systems: Integration into multi-stage question-answering systems where paragraph selection precedes detailed reading.
- Error Analysis: Studying the impact of paragraph selection accuracy on overall QA performance.
Limitations
It's important to note that this model assumes the benchmark-provided candidate boundary and is not suitable for open-domain retrieval, embedding search, or general reranking without specific validation. It is a legacy Llama 3 checkpoint, distinct from newer Llama 3.1 experiments.