research: SweRankEmbed-Small short-list vote (k=30 passes pre-registered bar on all 4 halves; k=100 fails dev-rest) by ParsaVictor · Pull Request #150 · ParsaVictor/code-context-engine

research: pre-register SweRankEmbed-Small as dense short-list embedder (bar: func@5 >= +0.02 on dev-fast, dev-rest, gym-a, gym-b; w=0.5, k=30)

ParsaVictor

and

claude

committed

79b257d

research: dense_shortlist --model swerank-embed-small (torch NomicBERT forward, CLS pooling, query prefix; jina default unchanged)

ParsaVictor

and

claude

committed

d47346d

research: dense_compare.py (engine / vote w / dense alone per half, secs) and paired.py vote-vs-vote mode

ParsaVictor

and

claude

committed

51f589b

research: pre-register k=100 SweRank short-list test (same bar, w=0.5) before running it

ParsaVictor

and

claude

committed

326724a

log: §8.19 SweRankEmbed-Small short-list vote clears the pre-registered bar on dev and SWE-Gym halves; +0.015 over jina, not significant

ParsaVictor

and

claude

committed

582fb83

research: fix holdout selection rule (k=100 only if it passes all four halves and pooled func@5 >= k=30) before any k=100 number

ParsaVictor

and

claude

committed

480e92f

log: §8.19 k=100 SweRank vote fails the bar on dev-rest (+0.006); holdout variant by the fixed rule = k=30, w=0.5

ParsaVictor

and

claude

committed

0630bd5

Navigation Menu

Repository navigation

research: SweRankEmbed-Small short-list vote (k=30 passes pre-registered bar on all 4 halves; k=100 fails dev-rest) - #150

research: SweRankEmbed-Small short-list vote (k=30 passes pre-registered bar on all 4 halves; k=100 fails dev-rest)#150
ParsaVictor merged 7 commits into
mainfrom
s19-swerank

Commits on Oct 17, 2026

research: pre-register SweRankEmbed-Small as dense short-list embedder (bar: func@5 >= +0.02 on dev-fast, dev-rest, gym-a, gym-b; w=0.5, k=30)

research: dense_shortlist --model swerank-embed-small (torch NomicBERT forward, CLS pooling, query prefix; jina default unchanged)

research: dense_compare.py (engine / vote w / dense alone per half, secs) and paired.py vote-vs-vote mode

research: pre-register k=100 SweRank short-list test (same bar, w=0.5) before running it

log: §8.19 SweRankEmbed-Small short-list vote clears the pre-registered bar on dev and SWE-Gym halves; +0.015 over jina, not significant

research: fix holdout selection rule (k=100 only if it passes all four halves and pooled func@5 >= k=30) before any k=100 number

log: §8.19 k=100 SweRank vote fails the bar on dev-rest (+0.006); holdout variant by the fixed rule = k=30, w=0.5