CrossMem-SQL IP&M Evaluation Data

Evaluation data and bank construction code supporting the manuscript"Teacher-Generated Schema Memory for Small Text-to-SQL Readers", submitted toInformation Processing & Management. The archive supports reproduction of the reported evaluation and statisticalanalyses, and documents how the teacher-generated demonstration banks wereconstructed. It contains the final selected demonstration banks, per-questionmodel predictions, retrieval records, per-row correctness and paired labels,derived statistics, frozen result summaries, dataset manifests, audit records,and the bank generation and selection code. A coverage ledger maps all 39reported results to their supporting artifacts and row-level evidence. Spider and BIRD are not redistributed. Obtain them from their originaldistributors; dataset-manifests/ records the versions, geometry and join keysneeded to align the released records with the benchmark rows. Exact regeneration of the banks is not claimed. The teachers are hosted modelswhose serving may change over time, so the frozen banks in banks/ remain theauthoritative artifacts for every reported result. The generation code carriesno endpoint, credential or model identifier; see generation-code/README.md. Verify the contents with `sha256sum -c CHECKSUMS.sha256`. MANIFEST.json is themachine-readable inventory. Licensing: the generated data are released under CC BY 4.0. The contents ofgeneration-code/ are released under the MIT licence (LICENSE-CODE.txt).Third-party benchmarks, model outputs subject to provider terms, and modelweights retain their original terms. Creators are anonymized for double-anonymized peer review and will be updatedon acceptance.