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
fix(research): let the bench harnesses take their model root from LAYA_MODELS
thread3dc7958dbf1602
diff --git a/research/README.md b/research/README.mdindex 9ce68d7c..7b8cbcf3 100644--- a/research/README.md+++ b/research/README.md@@ -32,6 +32,33 @@ the numbers quoted in the main README — nothing here is imported by the `laya`Everything runs with `USE_TF=0` — `transformers` probes for TensorFlow at import, and when TF isinstalled its abseil runtime can deadlock model construction on macOS/Python 3.9.+## Running the harnesses++The checkpoints come from `LAYA_MODELS` (default `~/laya_models`, a directory holding `laya/`,+`laya-multilingual/` and `laya-typed-decisions/`), so a checkout can point them at a local+`models/` copy:++```bash+# from the repository root+LAYA_MODELS=$PWD/models python research/scripts/bench_local.py --langs 10 --per-lang 60 --skip-b+LAYA_MODELS=$PWD/models BENCH_N=80 python research/scripts/bench_apps.py+LAYA_MODELS=$PWD/models python research/scripts/bench_latency.py+```++| option | effect |+|---|---|+| `bench_local.py --langs N` | caps the language list (sorted, so `10` is the first ten codes; `0` = all 51) |+| `bench_local.py --per-lang N` | cases per language (default 120) |+| `bench_local.py --skip-a` / `--skip-b` | run only the typed-decisions half, or only the MASSIVE half |+| `bench_apps.py` with `BENCH_N` | cases per application suite (default 400) |++`bench_local.py` and `bench_apps.py` write `research/*_benchmark_results.json`, which is+gitignored — the curated runs live in `results/`. Set `HF_HOME` inside the checkout to keep the+dataset cache local.++`bench_local.py` and `bench_latency.py` take the model root from `LAYA_MODELS` rather than assuming+`~/laya_models`, so the harnesses run against a checkout that keeps its weights under `models/`.+## Length batching[Local CPU measurements](results/length_batching_cpu_20260924.json) compare the originaldiff --git a/research/scripts/bench_latency.py b/research/scripts/bench_latency.pyindex 43dec911..b2207d65 100644--- a/research/scripts/bench_latency.py+++ b/research/scripts/bench_latency.py@@ -30,7 +30,7 @@ import laya # noqa: E402from laya.lang import analyse # noqa: E402from laya.router import Router # noqa: E402-ROOT = os.path.expanduser("~/laya_models")+ROOT = os.environ.get("LAYA_MODELS", os.path.expanduser("~/laya_models"))MODELS = {"english": os.path.join(ROOT, "laya"),"multilingual": os.path.join(ROOT, "laya-multilingual"),"typed-decisions": os.path.join(ROOT, "laya-typed-decisions")}diff --git a/research/scripts/bench_local.py b/research/scripts/bench_local.pyindex 3a3958fe..7babd5a4 100644--- a/research/scripts/bench_local.py+++ b/research/scripts/bench_local.py@@ -32,7 +32,7 @@ sys.path.insert(0, os.path.dirname(REPO))import laya # noqa: E402from laya.common import QTYPES, build_sequence, collate_items, render_options, temp_bucket # noqa: E402-ROOT = os.path.expanduser("~/laya_models")+ROOT = os.environ.get("LAYA_MODELS", os.path.expanduser("~/laya_models"))MODELS = {"english": os.path.join(ROOT, "laya"),"multilingual": os.path.join(ROOT, "laya-multilingual"),"typed-decisions": os.path.join(ROOT, "laya-typed-decisions")}