Emotion · lora (English): NoulXP package
The NoulXP package of biplov/emo-en-lora-proper, a Laya fine-tune made with System One Studio from aac6fef/laya-mlx. Its task: Which emotion does the writer express? (6 labels)
A NoulXP package runs on any NoulXP runtime (noulxp run, noulxp serve, the System One Engine) with no model-specific code: nothing in it executes.
Lineage
answerdotai/ModernBERT-large (Apache-2.0) → Laya (English) (Convai Innovations, Apache-2.0) → aac6fef/laya-mlx (MLX port, Apache-2.0) → biplov/emo-en-lora-proper (fine-tune, Apache-2.0) → this package.
What changed
- Format: ONNX graph, opset 23 (attention as ONNX's fused Attention operator), with the NoulXP files that describe the input, calibration and answers.
- Precision: none.
model.safetensorsis the published file byte for byte (float16 weights, computed in float32 as the model's own code computes them). - Nothing else.
noulxp/conformance.jsonlholds what the model's own code (laya0.3.5) answers to NoulXP's 52 requests on a CPU, and this package reproduces it.
Measurements (2026-10-06)
The reference is the model's own code (laya.load(...).system_one, float32, CPU), unrounded. The typed-decisions test split is LocalLLaMA/typed-decisions (400 requests, 2,000 questions). A package is published only if no answer changes there, ties included.
| Where | NoulXP's 52 requests | Typed-decisions test split (2,000 questions) | Speed |
|---|---|---|---|
| CPU (Intel(R) Xeon(R) Gold 6342 CPU @ 2.80GHz), onnxruntime 1.30.0 | pass: 52/52 cases within 0.01, max |Δp| 5.0e-05, 0 top answers changed | 0 of 2000 top answers changed (ties included), max |Δp| 1.2e-05, refusals that differ: 0 | p50 118.81 ms, p95 893.51 ms (one question per request, 4 threads) |
| NVIDIA A40, onnxruntime-gpu 1.30, exact | pass: 52/52 cases within 0.01, max |Δp| 5.0e-05, 0 top answers changed | 0 of 2000 top answers changed (ties included), max |Δp| 2.5e-05, refusals that differ: 0 | 56.1/s |
| NVIDIA A40, onnxruntime-gpu 1.30, fast | pass: 52/52 cases within 0.01, max |Δp| 0.0050, 0 top answers changed | 2 of 2000 top answers changed (ties included), max |Δp| 0.0036, refusals that differ: 0 | 82.76/s |
Its own question (Which emotion does the writer express? (6 labels)) on the 400 states of the test split: 0 of 400 answers changed, max |Δp| 4.2e-06.
Where it runs
On the CPU. The System One Engine's GPU image has onnxruntime-gpu 1.23.2, which cannot run opset 23's fused attention on CUDA, and NoulXP refuses such a graph there rather than run it on the CPU inside a CUDA session.
Use it
pip install systemonemodels
systemone run noulxp biplov/emo-en-lora-proper-noulxp --request request.json
Or with the files of this repository's noulxp/ folder and the reference runtime (pip install "noulxp[onnx]"):
noulxp check noulxp/ # replays the conformance file on this machine
noulxp run noulxp/ --request request.json # one request
noulxp serve noulxp/ # POST /v1/systemone
Licence
Apache-2.0, as the original. noulxp/LICENSE and noulxp/NOTICE say what the package contains. answerdotai/ModernBERT-large and Convai Innovations' Laya are Apache-2.0-licensed.