Cartly models
The weights the Cartly shopping search app runs on, for CPU inference with ONNX Runtime. Cartly runs locally on
your own computer; its setup script downloads these files and checks them against release/manifest.json in the
code repository.
| File | What it is |
|---|---|
models/onnx/laya/heads_fp32.onnx (+ .data) |
Cartly's decision layer: a Laya encoder fine-tuned with one classification head per question (intent, reference, brand role, visual intent, budget strictness, specificity, category). English and Hinglish |
models/laya/heads/ |
Its head configuration and tokenizer |
config/laya_calibration/heads.json |
Temperature per head, fitted on a held-out calibration split |
models/onnx/qwen3_emb_w8.onnx (+ .data) |
Qwen3-Embedding-0.6B, output cut to 512 dimensions, weight-only 8-bit, for query text |
models/onnx/siglip2_text_fp32.onnx (+ .data) |
SigLIP2's text tower, so a text query can be compared with product image vectors |
The tokenizers of the two encoders and SigLIP2's image tower are downloaded from their own repositories at pinned revisions; they are not copied here.
Training and evaluation
The Laya heads were fine-tuned on synthetic shopping conversations built from the catalog, and evaluated on a
hand-written test set (G1, about 330 English and Hinglish queries) that was never used for training. Numbers and
method: docs/05-laya-decision-layer.md and docs/09-evaluation-and-phases.md in the code repository.
Licence
Apache-2.0, as the base models. See NOTICE.