CLI Reference

fastcxt-simulate

usage: fastcxt-simulate [--scenario NAME] [--data-dir DIR] [options]

Simulate tree sequences for fastcxt training.

--scenario          Scenario name: species code (AnoGam, HomSap, ...) or
                    msprime scenario (constant, sawtooth, island)
--data-dir          Output directory for .trees files
--num-ts            Number of tree sequences (default: 1000)
--n-samples         Diploid individuals (default: 25)
--sequence-length   Segment length in bp (default: 1000000)
--mutation-rate     Override mutation rate
--recombination-rate  Override recombination rate
--genetic-map       stdpopsim genetic map name
--num-processes     Parallel workers (default: 8)

fastcxt-preprocess

usage: fastcxt-preprocess [--base-dir DIR] [options]

Preprocess tree sequences into training data.

--base-dir          Root directory with .trees files
--out-subdir        Output subdirectory name (default: processed)
--window-size       Base window size in bp (default: 2000)
--sequence-length   Expected sequence length (default: 1000000)
--num-pairs         Pairs to sample per tree sequence (default: 200)
--simplify-n        Simplify to first N samples (default: 0 = keep all)
--train-ratio       Train/test split ratio (default: 0.9)
--global-seed       Random seed (default: 12345)
--skip-existing     Skip already-processed files
--num-workers       Parallel workers (default: CPU count)
--extract-trees     Compute tree topology features
--max-samples       Pad tree features to this sample count (default: 0 = per-file).
                    Use when training on variable sample sizes to ensure
                    consistent feature dimensions (e.g. --max-samples 200).
--mutation-rate     Override mutation rate
--accessibility-mask  Path to .npz accessibility mask

fastcxt-train

usage: fastcxt-train [--model PRESET] [--dataset-path DIR] [options]

Train a fastcxt model.

--model             Model preset: small, base, large, base_trees
--dataset-path      Path to preprocessed data
--gpus              GPU device IDs (default: 0)
--epochs            Training epochs (default: 10)
--lr                Maximum learning rate (default: 3e-4)
--batch-size        Batch size (default: 128)
--grad-accum        Gradient accumulation steps (default: 4)
--workers           Data loading workers (default: 8)
--checkpoint        Resume from checkpoint
--log-dir           Logging directory

fastcxt-benchmark

usage: fastcxt-benchmark [--mode MODE] [options]

Run scaling benchmarks.

--mode              Modes: fastcxt_notree, fastcxt_tree, all
--sample-sizes      Sample sizes to benchmark (default: 5 10 25 50 100)
--device            Device (default: cuda:0)
--batch-size        Batch size (default: 64)
--output            JSON output path for results