SemiHS-Bench / data /STATISTICS.md
rajuptvs-hal's picture
Add benchmark data files
3fbd52f verified
|
Raw
History Blame Contribute Delete
2.36 kB

STATISTICS - eval split (900 records)

By source (tier1_source)

key count
BOL 318
catalog 582

By HS4

key count
2804 24
2805 4
2826 16
2844 1
2850 28
3707 29
3818 57
7104 5
7320 1
7326 1
7616 8
8443 1
8471 16
8473 68
8486 81
8504 81
8513 2
8514 22
8516 1
8517 11
8518 1
8523 34
8524 3
8526 2
8536 38
8538 3
8541 81
8542 81
8543 7
8544 12
9001 2
9013 1
9025 7
9027 6
9029 1
9030 81
9031 81
9405 2

By confidence tier

key count
high 525
low 3
medium 372

By segment

key count
material 164
equipment 113
metrology 176
component 413
end_product 34

By difficulty tag

718 of 900 eval records (79.8 %) sit on at least one confusable HS boundary; 84 carry more than one tag. A record is counted under every tag it carries, so the per-tag counts below sum to more than 718. Tags come in two groups (within-family sibling splits and cross-family frontiers); see construction/configs/boundary_tags.yaml for the deciding criteria, and the boundary_note field on each record for the human-readable rationale.

Group A — within-family sibling splits

tag count
8541_siblings 81
8542_ic_function 81
8486_process_stage 81
8504_power_splits 81
9030_measurement_splits 81
9031_inspection_splits 81
8536_connection_splits 38
3707_photochemical 29
2804_gas_purity 24
8471_adp_splits 16
9027_analysis_splits 6

Group B — cross-family frontiers

tag count
8541_vs_8542 55
storage_boundary 45
furnace_boundary 20
doped_vs_undoped 18
crystal_substrate_boundary 16
sensor_boundary 15
parts_attribution 13
amplifier_boundary 7
cable_vs_connector 6
machine_with_function 4
led_device_vs_luminaire 3
display_module_boundary 2
populated_board_boundary 0
process_vs_metrology 0

Two Group-B frontiers (populated_board_boundary, process_vs_metrology) have no eval records; both appear in the train split.