RankLLaMA
Stub. Created as a placeholder — expand with vault-sourced content.
Definition
RankLLaMA is a LLaMA-based pointwise reranker (from the RepLLaMA/RankLLaMA line, Tevatron) — a decoder LLM fine-tuned to score query-document relevance, representing the LLM-reranker end of the cost/quality spectrum.
Vault references (existing coverage)
- Learning to Rank — listed under “Neural LTR”
TODO
- Cite the source paper; contrast with prompt-based RankGPT (no fine-tuning).
Related
- Reranking · LLM as Judge · RankGPT · MonoT5 · Learning to Rank · Relational Transformer — RelativeDB’s 85M RT reranker is benchmarked as quality-comparable to RankLLaMA-7B at a fraction of the FLOPs
Articles
- Relational Reranking - Scoring Search Results with Structured Facts — positions RT against RankLLaMA-7B on a FLOPs-vs-NDCG@10 chart