Replication dataset for "Multi-Source Prior Pooling for Recorded Hard-Drive Failures: An Evaluation of Local Language Models"

This repository contains records from the local language-model data collection used in Multi-Source Prior Pooling for Recorded Hard-Drive Failures: An Evaluation of Local Language Models. The study evaluates whether elicited prior distributions improve predictions of recorded hard-drive failure counts beyond a historical covariate model using Backblaze Drive Stats. The collection covers six models accessed through LM Studio: Qwen3-14B, gpt-oss-20b, Mistral Small 3.2, Gemma 3 27B, Phi-4, and Llama 3.1 8B Instruct. It comprises 73,800 planned response slots across 20 development cohorts and 21 evaluation cohorts, with ten batches of 30 responses per model and cohort. The repository includes: protocol.json: The recorded collection protocol and request configuration. responses.jsonl: The exported response records, stored as one JSON object per line. seed_manifest.json: The request-to-seed assignments used to organize generation and response retries. summary.json: The collection summary, including completion counts, response-validation results, retries, and comparator fallbacks. These files document the collection underlying the reported analyses. They should be interpreted alongside the manuscript’s definitions of response validity, numerical fitting, and fallback handling. Model identifiers and random seeds alone do not guarantee identical outputs across inference environments. Reproducing the complete analysis also requires the cohort data, historical priors, analysis code, and relevant software configuration.