{"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=GSM8K","json_url":"https://assignee.net/benchmarks/evidence.json?model=Llama-2&bench=GSM8K","model":"Llama-2","benchmark":"GSM8K","source_count":7,"source_coverage":{"record_count":7,"distinct_source_count":7,"coverage_level":"BROAD","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":7,"missing_source_url_count":0,"domains":["arxiv.org","doi.org"],"year_min":2023,"year_max":2024,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":28.8,"max_score_pct":72.0},"spread_pp":43.2,"severity":"HIGH","entries":[{"model":"Llama-2","benchmark":"GSM8K","score_pct":72.0,"source_title":"Achieving >97% on GSM8K: Deeply Understanding the Problems Makes LLMs Better Solvers for Math Word Problems","source_url":"http://arxiv.org/abs/2404.14963v5","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"GSM8K","score_pct":66.6,"source_title":"Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations","source_url":"http://arxiv.org/abs/2312.08935v3","source_domain":"arxiv.org","year":2023},{"model":"Llama-2","benchmark":"GSM8K","score_pct":53.3,"source_title":"SmallToLarge (S2L): Scalable Data Selection for Fine-tuning Large Language Models by Summarizing Training Trajectories of Small Models","source_url":"http://arxiv.org/abs/2403.07384v2","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"GSM8K","score_pct":52.08,"source_title":"Understanding Reasoning in Chain-of-Thought from the Hopfieldian View","source_url":"http://arxiv.org/abs/2410.03595v1","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"GSM8K","score_pct":40.0,"source_title":"Making Large Language Models Better Reasoners with Alignment","source_url":"http://arxiv.org/abs/2309.02144v1","source_domain":"arxiv.org","year":2023},{"model":"Llama-2","benchmark":"GSM8K","score_pct":35.0,"source_title":"Learning From Failure: Integrating Negative Examples when Fine-tuning Large Language Models as Agents","source_url":"http://arxiv.org/abs/2402.11651v2","source_domain":"arxiv.org","year":2024},{"model":"Llama-2","benchmark":"GSM8K","score_pct":28.8,"source_title":"TRACE: A Comprehensive Benchmark for Continual Learning in Large Language Models","source_url":"https://doi.org/10.48550/arxiv.2310.06762","source_domain":"doi.org","year":2023}],"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."]}