bactrianus/HotpotQA-Reader-Llama-3-8B-Instruct

Model Overview

The bactrianus/HotpotQA-Reader-Llama-3-8B-Instruct is a specialized Llama 3 8B Instruct model, fine-tuned by Bactrianus to function as a HotpotQA-focused reader. Its primary role is to process multi-hop questions alongside pre-selected evidence and generate concise answers. This model is a historical artifact, distinct from the Llama 3.1 checkpoints discussed in the revised Bactrainus manuscript.

Key Capabilities

• Direct Answer Generation: Receives a multi-hop question and supplied evidence, then produces a concise answer.
• Evidence-Based Reading: Optimized for generating answers directly from provided supporting sentences and their titles.
• Research on Answer Generation: Ideal for studying the reader stage independently in complex question answering systems.
• Integration into Pipelines: Designed to be the final component in a fixed-candidate selector-reader pipeline, where evidence selection is handled upstream.

Intended Use Cases

• Research: Conducting research on answer generation from compact multi-document evidence.
• Ablation Studies: Performing reader-stage ablations within the English HotpotQA distractor setting.
• System Integration: Integrating as a reader component behind a paragraph/sentence selector that provides validated evidence.

Limitations

• Specialized Scope: Primarily designed for English HotpotQA-style inputs; not validated as a general-purpose reader.
• Evidence Dependency: Accuracy is highly dependent on the quality of supplied evidence; missing facts cannot be reliably recovered.
• Context Noise: Extra paragraphs can introduce substantial noise, potentially affecting answer quality.
• Output Variability: Generated answers may occasionally be unsupported, malformed, or more verbose than expected.
• Data Biases: Inherits temporal, cultural, and coverage biases from its Wikipedia-derived training data.