Archived Model Outputs for POS-Signature Demonstration Retrieval for In-Context Dependency Parsing

This artifact contains the archived model responses and scored evaluation records for “POS-Signature Demonstration Retrieval for In-Context Dependency Parsing.” The study evaluates openai/gpt-oss-120b and Qwen/Qwen2.5-72B-Instruct on all 2,077 sentences of the Universal Dependencies English-EWT test split under six prompting conditions: Zero-Shot, Fixed Few-Shot, POS-Signature Few-Shot, Semantic Few-Shot, Chain-of-Thought, and Critique-Refine. The deposit contains 28 response, evaluation, and pre-recovery JSON archives totaling 749,638,691 uncompressed bytes. It also includes SHA-256 manifests, prompt-provenance records, the authoritative paper-result object, and the applicable EWT license and notices. The archives support offline reproduction of corpus UAS/LAS, paired sentence-bootstrap comparisons, reliability analyses, recovery sensitivity, relation and length diagnostics, critique churn, and prompt-provenance verification without additional model API calls. Missing or unscorable predictions remain in the common 2,077-sentence and 21,998-token evaluation denominator. Source code and reproduction instructions are available in the associated Git repository.