code wiki / _hdl_build / nx_drbench.nx
nx_drbench.nx
buildroot/runtime/_hdl_build/nx_drbench.nx
about
nx_drbench.nx -- HotpotQA distractor val[0..99]: the FIRST public-benchmark measurement of the Nishi
researcher (census axis [BEN]). Protocol = published distractor eval (10 paras: 2 gold + 8 distractors);
retrieve + extract; scored by SQuAD word-F1/EM (nx_qa_score_lib, integer permille) + gold-para recall@2.
MODES: mode0 single-hop lexical (IDF-weighted para retrieval + span-heuristic extraction); mode1 mechanical
multi-hop (hop-1 best para/sentence -> bridge entities -> expanded second retrieval). TEETH: T1 parse
integrity, T2 gold-title resolution, T3 rotated-gold NEG-CONTROL (~0), T4 retrieval floor.
MEASURED 2026-07-07 (100-row p0): MODE0 F1=172 EM=13 r@2=640 / MODE1 F1=155 EM=11 r@2=660 -- the honest
mechanical ceiling (bridge-not-answer-target failure mode; the climb is census R1/R3).
DEBT EATEN 2026-07-07: the ~1100-line engine that lived here moved to nx_qabench_engine.nx (shared with
nx_qabench's multi-benchmark ladder) -- this file is now IMPORT + main only, zero duplication, one fix point.
ctx slots + engine docs live in nx_qabench_engine.nx. expect_exit: 0 license_tier: ORIGINAL
dependencies 1 imports · 0 importers
imports: nx_qabench_engine.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| 13 | const K_MAGIC_1024: i64 = 1024 |
| 14 | const K_MAGIC_1572864: i64 = 1572864 |
| 15 | const K_MAGIC_4096: i64 = 4096 |
| 16 | const K_MAGIC_2048: i64 = 2048 |
| 17 | const K_MAGIC_8192: i64 = 8192 |
| 18 | const K_MAGIC_262144: i64 = 262144 |
| 19 | const K_MAGIC_131072: i64 = 131072 |
| 20 | const K_MAGIC_32768: i64 = 32768 |
| 21 | const K_MAGIC_4000: i64 = 4000 |
| 22 | const K_MAGIC_2000: i64 = 2000 |
| 23 | const K_MAGIC_8000: i64 = 8000 |
| 24 | const K_MAGIC_1000000: i64 = 1000000 |
functions
| 26 | func main() -> i64 |