bactrianus/HotpotQA-Reader-CoT-Llama-3-8B-Instruct
The bactrianus/HotpotQA-Reader-CoT-Llama-3-8B-Instruct is a Llama 3 8B Instruct-based causal language model fine-tuned by bactrianus for generating natural-language rationales and answers from provided evidence. This model is specifically adapted for HotpotQA-style inputs, focusing on producing an intermediate rationale before the final answer. Its primary use case is studying rationale supervision in complex question answering and qualitative inspection of evidence-conditioned answer paths.
Loading preview...
Model Overview
This model, bactrianus/HotpotQA-Reader-CoT-Llama-3-8B-Instruct, is a fine-tuned variant of meta-llama/Meta-Llama-3-8B-Instruct. Developed by bactrianus, its core function is to generate an intermediate natural-language rationale followed by a final answer, based on supplied evidence for HotpotQA-style questions. The rationales are a form of process supervision for task adaptation, not hidden traces from the base model.
Key Capabilities and Intended Use
- Rationale and Answer Generation: Produces both a natural-language rationale and a final answer from a given question and selected evidence.
- HotpotQA Research: Designed for studying natural-language rationale supervision within the HotpotQA dataset context.
- Qualitative Inspection: Useful for qualitatively inspecting how evidence conditions the answer path.
- Reader-Stage Comparisons: Can be used in scenarios where evidence is supplied independently for reader-stage comparisons.
Training and Lineage
The model is derived from Meta Llama 3 8B Instruct and HotpotQA-based reader/rationale supervision datasets. Relevant datasets include cot-reader-sft for complete deterministic views and teacher-cot-llama31-8b-sft and teacher-cot-llama31-70b-sft for Llama 3.1-labelled rationales, though this specific model is a legacy Llama 3 checkpoint.
Limitations
- Generated rationales may be post-hoc, incomplete, contradictory, or unsupported.
- Specialized for English HotpotQA-style inputs.
- Evidence omissions can propagate to both rationale and answer.
- Not intended for open-domain retrieval, safety-critical decisions, or factual verification.