{"schema":"https://assignee.net/schemas/benchmark-evidence-v1","schema_version":"1.0","contract_version":"benchmark-evidence-v1.0","contract_updated":"2026-06-01","schema_documentation":"https://assignee.net/schemas","changelog_url":"https://assignee.net/changelog","publisher":{"name":"Assignee Research","url":"https://assignee.net"},"html_url":"https://assignee.net/benchmarks/evidence?model=Llama-2&bench=MMLU","json_url":"https://assignee.net/benchmarks/evidence.json?model=Llama-2&bench=MMLU","model":"Llama-2","benchmark":"MMLU","source_count":4,"source_coverage":{"record_count":4,"distinct_source_count":4,"coverage_level":"MODERATE","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":4,"missing_source_url_count":0,"domains":["arxiv.org","openalex.org"],"year_min":2024,"year_max":2026,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":37.8,"max_score_pct":46.7},"spread_pp":8.9,"severity":"MEDIUM","entries":[{"model":"Llama-2","benchmark":"MMLU","score_pct":46.7,"source_title":"Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs","source_url":"http://arxiv.org/abs/2407.15549v3","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"MMLU","score_pct":44.8,"source_title":"Data Engineering for Scaling Language Models to 128K Context","source_url":"http://arxiv.org/abs/2402.10171v1","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"MMLU","score_pct":43.9,"source_title":"Rotated Robustness: A Training-Free Defense against Bit-Flip Attacks on Large Language Models","source_url":"https://openalex.org/W7139145179","source_domain":"openalex.org","year":2026},{"model":"Llama-2","benchmark":"MMLU","score_pct":37.8,"source_title":"Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes","source_url":"http://arxiv.org/abs/2403.00867v3","source_domain":"arxiv.org","year":2024}],"interpretation":"This record groups score claims extracted from papers for the same model and benchmark label. A nonzero spread means the public literature reports different values for this cluster.","limitations":["Differences are not automatically errors.","Reported values may differ because of prompts, dataset versions, evaluation protocols, scoring rules, preprocessing, fine-tuning, or reporting conventions.","Source papers remain authoritative for their own claims."]}