A 0-100 scale didn't stop gemma3:12b's tied scores, but breaking ties with the model's digit probabilities put the right article first on 11 of 12 questions.
Badges
Badges come from GitHub once the author claims this feed.
Recent posts
every link opens the author’s post-
Broke the Re-Ranker's Ties With Its Own Probabilities: 11 of 12 First, at 95 Seconds a Question
-
Re-Ranking the Top 20 With Local Models: Only the 12B Fixed the Miss, and It Demoted Two Right Answers
Re-scored the top 20 chunks for 12 questions with 1B, 3B and 12B local models: only the 12B fixed the clear miss, it demoted two right answers, and the 3B scored below no re-ranking on top-1.
-
Swept the Per-Article Cap From 1 to 3: The Gains Run Out Before the Costs Do
Ran the 12-question eval at caps of none, 1, 2 and 3 chunks per article: the hit-rate gains came from two questions and vanished by cap 3, while depth kept shrinking.
-
Scored the Index Against Known Answers: 10 of 12, and What the Cap Really Costs
Scored a 741-chunk local index against 12 questions with known right articles: 10 hit first place, and the per-document cap pulled both misses into the top five at the cost of depth.
-
Multi-AZ Isn't a DR Plan: Zone Outage on OpenShift and AWS
A zone outage on a ROSA HCP cluster showed why multi-AZ isn't disaster recovery: stateless services recovered in 25 minutes, stateful ones took 3h15m.
-
Tested a Per-Document Cap at Real Scale — and Found I Had No Ground Truth
Measured chunking bias against a 741-chunk collection and tried a per-document result cap.