bactrianus/HotpotQA-Sentence-Retriever-Llama-3-8B-Instruct

Model Overview

The bactrianus/HotpotQA-Sentence-Retriever-Llama-3-8B-Instruct is a specialized causal language model, derived from meta-llama/Meta-Llama-3-8B-Instruct. Developed by bactrianus, this model's primary function is to identify and extract supporting sentences from given paragraphs in response to multi-hop questions, specifically within the context of HotpotQA-style inputs.

Key Capabilities

• Supporting-Sentence Selection: Receives a multi-hop question and selected paragraphs, then outputs the exact title and zero-based sentence index of each predicted supporting fact.
• Precise Fact Interface: Designed to feed a reader through a compact facts interface, providing granular, sentence-level information.
• HotpotQA Optimization: Fine-tuned using HotpotQA supporting-fact annotations, making it highly effective for tasks requiring detailed evidence extraction from provided text.
• Input/Output Contract: Requires exact paragraph titles, sentence order, and zero-based indices for input, and outputs deduplicated (title, sentence_index) pairs, with validation against the input.

Intended Use Cases

• HotpotQA-style Supporting-Fact Selection: Ideal for tasks that involve identifying specific sentences as evidence for multi-hop questions.
• Sentence-Level Cascading: Suitable for integration into systems that require a paragraph-to-sentence cascade for detailed information retrieval.
• Analysis of Exact Set Prediction: Useful for research and analysis concerning exact set prediction and error propagation in fact extraction.

Limitations

It's important to note that this model's performance is dependent on the quality of the initial paragraph selection. It is not designed for searching outside supplied paragraphs, general sentence ranking in unrelated domains, or as a general evidence-verification model without external validation. The model assumes HotpotQA-style title and sentence boundaries.