muat: preprocessed data and reproducibility checkpoints

Preprocessed input files, trained model checkpoints, and training/evaluation sample splits supporting the reproducibility experiments in "muat: portable transformer-based method for tumour classification and representation learning from somatic variants" (Sanjaya & Pitkänen). Preprocessed data pcawg_snvmnvindelsv_muat.tar.gz — preprocessed open-access PCAWG whole-genome cohort, one .muat.tsv file per sample (1,794 samples, 17 tumour types, GRCh37/hg19, SNV+MNV+indel+SV), ready for direct use with muat's --no-preprocess flag. pubhg38_tcga_muat.tar.gz — preprocessed public GDC TCGA cohort, one .muat.tsv file per sample (502 samples: 413 training / 89 validation, 6 tumour types, native GRCh38/hg38), plus the corpus-level dictionaries and exact train/validation splits used, ready for direct use with muat's --no-preprocess flag. Training/evaluation sample splits d_matrix_train_val_split.zip — the exact train/validation split (1,434 training / 360 validation samples; sample path into pcawg_snvmnvindelsv_muat.tar.gz, tumour-type label, class index) used to train and evaluate checkpoints ckpt1.zip and ckpt2.zip. a_b_matrix_train_val_split.zip — the 10-fold cross-validation split (sample ID, tumour-type label, class index per fold) defining the PCAWG controlled-access cohort's train/validation partitions used in the a0 (CSC) and b0 (Genomics England) reproducibility experiments. Sample IDs and tumour-type labels only — no mutation calls or other genomic data are included. Access to the underlying PCAWG controlled-access mutation data requires separate ICGC DACO authorization. Checkpoints trained in this study ckpt1.zip — 10 independently trained checkpoints (GPU, PCAWG, SNV+MNV+indel+SV). ckpt2.zip — 10 independently trained checkpoints (CPU, PCAWG, SNV+MNV+indel+SV). ckpt3.zip — 10 independently trained checkpoints (GPU, TCGA WES, native GRCh38/hg38). Original published checkpoints (Sanjaya et al. 2023), mirrored for a persistent, checksummed DOI benchmark_wes.zip — original MuAt whole-exome (TCGA) pretrained checkpoints, 10-fold cross-validation ensemble. Mirrored from the original Hugging Face release (huggingface.co/primasanjaya/muat-checkpoint); used by muat's predict-ensemble command. benchmark_wgs_snv.zip, benchmark_wgs_snv+mnv.zip, benchmark_wgs_snv+mnv+indel.zip, benchmark_wgs_snv+mnv+indel+svmei.zip, benchmark_wgs_snv+mnv+indel+svmei+neg.zip — original MuAt whole-genome (PCAWG) pretrained checkpoints, motif-only architecture, 10-fold cross-validation ensembles, one archive per mutation-type configuration (50 checkpoints total). benchmark_wgs_posges_<mutation-type>_pos.zip / _posges.zip (10 archives) — the same original PCAWG pretrained checkpoints, with position (+pos) and position+genic/exonic/strand (+pos+ges) feature variants, one archive per mutation-type/feature-variant combination (100 checkpoints total). muat_katainen.zip -- checkpoint to reproduce muat on Finnish CRC (Katainen et al, Nature Genetics 2015) Together, the benchmark_wgs_*.zip and benchmark_wgs_posges_*.zip archives provide all 150 original PCAWG checkpoints (5 mutation-type configurations × 3 feature-set variants × 10 cross-validation folds) used as the baseline in this study's PCAWG controlled-access reproducibility (a0) and Genomics England portability (b0) experiments. ckpt1.zip, ckpt2.zip, and ckpt3.zip each include a MANIFEST.tsv (per-repeat SHA256 provenance) and a README.md. See the associated publication's Supplementary Data for exact download/verification commands.