bactrianus/HotpotQA-Reader-Llama-3-70B-Instruct

TEXT GENERATIONPricing:Input $3.5 / Cached $0.7 / Output $8.3Concurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Aug 3, 2024License:llama3Architecture:Transformer Featherless Exclusive Cold

The bactrianus/HotpotQA-Reader-Llama-3-70B-Instruct is a Llama 3 70B Instruct-based causal language model developed by bactrianus, specifically fine-tuned as a large direct answer reader. It excels at generating concise answers from a given question and supplied evidence, making it suitable for research into evidence-conditioned multi-hop answer generation. This model is specialized for HotpotQA-style tasks, focusing on synthesizing information from provided contexts rather than performing retrieval.

Loading preview...

Model Overview

This model, bactrianus/HotpotQA-Reader-Llama-3-70B-Instruct, is a specialized Llama 3 70B Instruct-based causal language model. It functions as a large direct answer reader, designed to generate concise answers based on a given question and explicitly supplied evidence. It does not perform passage selection or external corpus searches, relying entirely on the provided context.

Key Capabilities

  • Evidence-conditioned Answer Generation: Generates answers by synthesizing information from provided evidence.
  • HotpotQA Specialization: Optimized for tasks similar to HotpotQA, focusing on multi-hop question answering where supporting facts are supplied.
  • Research Tool: Primarily intended for research into capacity effects in evidence-conditioned multi-hop answer generation and reader-stage diagnostics.

Intended Use Cases

  • Research: Studying how large language models process and synthesize information from provided evidence for complex question answering.
  • Modular Pipelines: Serving as the final-stage generation component in pipelines where an upstream component selects and provides relevant evidence.
  • HotpotQA Diagnostics: Validating reader-stage performance with HotpotQA supporting facts.

Limitations

  • Specialized: Optimized for English HotpotQA-style evidence-conditioned answering; performance outside this domain may vary.
  • No Retrieval: Cannot perform open-domain retrieval or web search; relies entirely on supplied evidence.
  • Evidence Dependence: Cannot reliably repair missing or insufficient evidence, and full paragraphs or distractors can reduce answer quality.
  • Historical Artifact: This is a Llama 3 artifact; the revised Bactrainus manuscript uses Llama 3.1 70B Instruct, and results are not interchangeable.