Datasets:

neuralbioinfo
/

all-nr-plasmid-training-public

Parquet error: Scan size limit exceeded: attempted to read 302378734 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files

Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

All-NR Clean Plasmid 50Mbp 1:4 Top50-Switchable Dataset `actL4000`

This public all-NR plasmid-vs-host training dataset is freshly sampled from the clean NR profile skani_minaf80_ani99_afsym97p5_affull99_afpart95. It is not derived from a curr1 dataset. Plasmid sampling uses a 50Mbp host-genus budget and the strict training_clean config targets four clean host negatives per sampleable plasmid-positive segment after minimap2 filtering.

This upload was produced with config group switchable_top50_pool_only. In the default training_clean_only group, only the strict training dataset is published as a Hugging Face config. Summary and audit tables are kept locally beside the generated dataset rather than uploaded as public configs.

Generated: 2026-06-15 13:30:41 CEST

Repository: neuralbioinfo/all-nr-plasmid-training-public

config	split	rows
allnr_nr97p5_l200k_tsh50m_up5_down0p25_hc2p5_mm_p1h4_top50tag_actL4000_training_switchable_top50_pool	train	12563690

Config Semantics

config	description
allnr_nr97p5_l200k_tsh50m_up5_down0p25_hc2p5_mm_p1h4_top50tag_actL4000_training_switchable_top50_pool	Switchable top50 host-hit driver pool. It contains top50-driver positives and recoverable host rows; filter and rebalance after choosing top50 behavior.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "neuralbioinfo/all-nr-plasmid-training-public",
    name="allnr_nr97p5_l200k_tsh50m_up5_down0p25_hc2p5_mm_p1h4_top50tag_actL4000_training_switchable_top50_pool",
    split="train",
)

For top50-driver ablation with training_switchable_top50_pool, drop positive rows where top50_host_hit_driver_positive=True, keep host rows where filtered_by_non_top50_plasmid=False, and then sample the requested host:plasmid ratio for training.

Local Summaries Not Uploaded As Configs

The following local outputs are retained under /project/c_evolm/training_datasets/phage_db/plasmids/ncbi/TrainingSamples/plasmid_allnr_nr97p5_l200k_tsh50m_up5_down0p25_hc2p5_mm_p1h4_top50tag for audit and debugging:

manifests/dataset_manifest.json
checks/minimap_filtering_summary.*
checks/final_dataset_summary.feather
checks/genus_balance_summary.feather
checks/top50_host_hit_driver_plasmids.feather
checks/top50_switchable_summary.feather
hf_upload/upload_summary.json

Key Counts

count	value
`n_plasmid_segments`	`2441012`
`n_host_segments_candidate`	`24193650`
`n_host_segments_minimap_aligned`	`14300940`
`n_host_segments_filtered_by_plasmid`	`1146414`
`n_host_segments_clean_selected`	`9707157`
`n_training_clean_rows`	`12148169`
`n_final_segmentdb_rows`	`13294583`
`n_long_suspicious_plasmids_excluded`	`6241`
`top50_host_hit_driver_count`	`50`
`n_training_switchable_top50_pool_rows`	`12563690`
`n_recoverable_if_drop_top50_hosts`	`415521`

Sampling Parameters

parameter	value
`target_genus_bp`	`50000000`
`max_upsampling_rate`	`5.0`
`min_downsampling_rate`	`0.25`
`clean_host_ratio`	`4`
`host_candidate_compensation_rate`	`2.5`
`minimap_qcov_threshold`	`0.95`
`minimap_min_pid`	`0.95`
`minimap_min_aligned_bp`	`3500`
`top50_hq_qcov`	`0.95`
`top50_hq_pid`	`0.95`
`top50_hq_min_cov_bp`	`3500`
`top50_hq_max_aln_to_query_ratio`	`1.1`