Plumb-4B (crh225, JevK5 v0.2 + LoRA)

Availability and evidence

Openness

Code and weights marked open in the published row

License note

Apache-2.0 (weights and code; NOTICE credits JevK5, Qwen and SemIf); base Qwen3.5-4B Apache-2.0

Cost evidence

estimated; the board’s row disclosure contains the published basis.

Endpoint condition

our evaluator-owned Lium GPU pod (H100 80 GB (evaluator-owned Lium pod)), offline read-only container, author's model served on loopback

Note on this row

Crh225/plumb-4b @ 55de037801a8a9b9de3db5c0e16cef86210c2186: merged bf16 weights, LoRA r16 on the attention projections of JevK5 v0.2 (itself Qwen3.5-4B), served by the author's documented command, jevk5 v0.2.0's own jevk5-serve (github.com/allebee/jevk5, Apache-2.0), unmodified. One forward pass per decision: softmax over the declared options' answer-letter logits at the last position divided by the package's own T = 2.07 (jevk5_config.json), up to 16 options, 0 generated tokens; inputs over 16,384 tokens are refused rather than truncated. Disclosed by the author: no JevBench item was trained or tuned on and every checkpoint and temperature choice came from his own held-out sets, but the 231 public items were scored after each of five training rounds and that aggregate feedback shaped the recipe (hard mining, long documents, document length) — development against the public distribution, stated plainly. His own overlap audit dropped 9 training items sharing more than two 8-word spans with a public item. Independent top-five gate (25 Sep 2026): LEGIT — cost basis (same as its JevK5 parent), public-versus-sealed gap against like-for-like rows, an 8-gram screen of its released training data (no row shares more than two 8-grams with a public item) and an exact composite recomputation. Offline self-hosted inference of all 842 decisions (534 frozen v1.2 + 308 sealed v1.4) in a network-disabled, read-only container on an evaluator-owned Lium H100 80 GB (Hopper) pod, through JevBench's unchanged typesafe adapter against the author's own server on loopback; no operator endpoint; no golds were exposed. Latency is the serial request wall time on the standard+judge items with the self-hosted adjustment. Cost is a labelled estimate at the EmpirioLabs Qwen3.5-4B public pay-as-you-go list price ($0.04/M input, no output generated), applied to this system's measured tokens; it is not a GPU bill. 25 Sep 2026 pre-release correction: retired DeepInfra $0.03/M input reference replaced by bookable EmpirioLabs $0.04/M exact-base input reference; draft score 67.09 to 65.84.

Published source

https://github.com/crh225/plumb

From the public v1.4.2.2 aggregate. Scores and ranks can change when a new release is published.