Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
ci(java): one parity cell at a time, because this change added HF download pressure
Adeshina Lasisi0e91bb45fc4f
diff --git a/.github/workflows/java.yml b/.github/workflows/java.ymlindex 9eed1977..e8a9617e 100644--- a/.github/workflows/java.yml+++ b/.github/workflows/java.yml@@ -394,6 +394,22 @@ jobs:timeout-minutes: ${{ matrix.timeout }}strategy:fail-fast: false+ # One cell at a time, because every cell that needs weights downloads them from the same+ # Hugging Face repo UNAUTHENTICATED, and the hub rate-limits that. This matrix grew a third+ # cell in this change, and `dotnet.yml`'s parity matrix runs four of its own against the+ # same repo in the same workflow run, so the peak was seven concurrent unauthenticated+ # fetches of one repo. Under that, `hf-xet` returned a TRUNCATED file without raising: the+ # 34 MB multilingual `tokenizer.json` arrived empty, the download step reported success, and+ # the failure surfaced ninety seconds later as `EOF while parsing a value at line 1+ # column 0` from inside a tokenizer constructor. Reproduced twice on this branch while the+ # same job passed on a branch that adds no cell.+ #+ # Serialising costs wall-clock on an advisory lane and removes the pressure this change+ # added. It is a MITIGATION, not the fix: the real defect is that nothing verifies a+ # download before using it, and the real fix is an `HF_TOKEN` for the higher rate limit or+ # a size/digest check at the fetch. Both live outside this directory, so they are not+ # bundled here.+ max-parallel: 1matrix:include:# Tokenizers and configs only -- no weights, nothing traced. The three families here