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

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Aug 3, 2024License:llama3Architecture:Transformer Featherless Exclusive Cold

The bactrianus/HotpotQA-OneStep-Retriever-Llama-3-8B-Instruct is an 8 billion parameter causal language model developed by bactrianus, based on Meta Llama 3. This model is specifically designed for joint supporting-fact and answer generation within the HotpotQA framework, operating over supplied candidate paragraphs. It excels at fixed-candidate joint evidence/answer generation research and comparison with modular systems, coupling evidence and answer errors in a single output.

Loading preview...

Model Overview

The bactrianus/HotpotQA-OneStep-Retriever-Llama-3-8B-Instruct is an 8 billion parameter causal language model built upon meta-llama/Meta-Llama-3-8B-Instruct. Developed by bactrianus, this model is specifically engineered for the HotpotQA dataset, focusing on joint supporting-fact and answer generation from a set of provided candidate paragraphs.

Key Capabilities and Features

  • Joint Evidence and Answer Generation: Unlike traditional retriever-reader systems, this model directly produces both supporting facts and the final answer in a single generation step, coupling potential errors in one output.
  • Fixed-Candidate Processing: It operates exclusively on candidate paragraphs supplied with each HotpotQA example, making it suitable for research into fixed-candidate question answering.
  • Research Tool: Intended for research purposes, particularly for comparing with modular paragraph, sentence, and reader systems, and analyzing parser/output-factorization failures.
  • Llama 3 Base: Leverages the architecture and capabilities of the Meta Llama 3 8B Instruct model.

Intended Use Cases

  • Fixed-candidate joint evidence/answer generation research.
  • Comparison with separately trained paragraph, sentence, and reader modules.
  • Analysis of parser and output-factorization failures.

Limitations

  • Evidence and answer errors are inherently coupled.
  • Not designed for corpus-scale retrieval or web search.
  • Sensitive to long candidate sets, potentially leading to distractor sensitivity and truncation.
  • Requires strict validation of generated titles, indices, and answers.

Finetunes
1 models