Ariadne-Laya-Support

Routes customer requests to 27 support intents. This 4.2 MB specialist interface runs on a shared frozen Laya base. Switching between compatible Ariadne specialists replaces about 1 million parameters, while the 421 million parameter base stays in memory. Each specialist uses its own small interface.

Use

Install the included Python wheel from this downloaded model folder:

pip install ./ariadne_specialists-0.2.0a1-py3-none-any.whl

import ariadne
from ariadne.specialists import Support

model = ariadne.load_specialist(Support, model=".")
result = model("Where is my refund?")
print(result.label, result.score)

The base downloads automatically and is cached. Pass device="cpu" or device="cuda:1" to choose a device. A list of texts returns a list of predictions. Scores have not been recalibrated for this task.

Load from Hugging Face

After installing the included wheel, you can load this repository directly:

import ariadne
from ariadne.specialists import Support

model = ariadne.load_specialist(Support, model="GoatHerder/Ariadne-Laya-Support")

Use the explicit model= argument with this preview wheel. The interface and pinned base are downloaded automatically and cached. Pass revision="<commit hash>" to pin a particular interface version.

Interface

The interface is a 1,024 × 1,024 linear projection plus a 1,024-element bias: 1,049,600 trainable parameters. It sits after the base's native embeddings and before encoder block 0. It starts as the identity; training updates only this projection. The shared base has 421,293,827 parameters. Compatible specialists share one resident base in the same Python process and on the same device.

Results

All local results below use the same 2,166 requests and label task.

Model	Accuracy	Macro-F1
Base Laya	97.14%	97.28%
Ariadne Support	99.77%	99.73%
learn-abc/magicSupport-intent-classifier	99.95%	99.96%
TF-IDF + logistic regression	99.22%	99.17%

The public fine-tune's training overlap with these cases is unknown, so this table does not establish a common unseen-test ranking. Full results and existing-task retention are in metrics.json.

Training and scope

Trained on bitext/Bitext-customer-support-llm-chatbot-training-dataset, revision 430d1a89bd93bd1fa23c16f29dd53e73f0087443. One seed (0); epoch 3 selected by validation loss. LR 1e-4; only the 1,049,600 interface parameters were trained. The base embeddings, 28 encoder blocks and decision heads stayed frozen.

English, single-topic inputs and the labels in task.json. This source-corpus benchmark is easy and does not establish accuracy on real customer traffic. Unsupported or ambiguous requests still receive a label. Source datasets retain their own licences.

Base revision: 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851. Independent adaptation; no affiliation with the original Laya authors.