MediaPipe Decision Maker β€” Laya Multilingual (256 & 512 Token)

Multilingual Laya (149M mmBERT-base) bidirectional cross-encoder bundles (laya_s256.task, 678.5 MB; laya_s512.task, 679.4 MB) for the MediaPipe Decision Maker (DecisionMaker) task, evaluating structured categorical choices, boolean checks, and ordinal rubrics in a single zero-decoding forward pass.

Documentation: MediaPipe Decision Maker Docs

Samples: Android Β· iOS Β· Python Colab Β· Web Β· Flutter

Try live: Web Demo

Derived from convaiinnovations/laya (mmBERT-base encoder, 100+ languages supported upstream; English and Japanese validated in litert-community/Laya-Multilingual-LiteRT) and packaged with FP16-weight (wfp16) 256- and 512-token graphs.

Inside DecisionMaker, up to β‰ˆ8 compact options fit alongside the context inside a single cross-attention forward pass (β‰ˆ4Γ— faster than pairwise evaluation), with larger option sets falling back to multi-pass evaluation.

Files

File Precision Size SHA-256
laya_s256.task β€” 678.48 MB (678,477,844 B) 8cb730b89ef99cb5c98cdbc7936de79deab77623ce25dc0031eec9e25c64c24e
laya_s512.task β€” 679.40 MB (679,395,348 B) e9f1693861c9a79a541d6f716648142ee844fd2e7119c9ed48d166f4b85a15df

Companion Files in Repo Root

File Role Size SHA-256
LICENSE Apache-2.0 license text (from Laya-Multilingual-LiteRT) 10.2 KB (10,173 B) a6cba85bc92e0cff7a450b1d873c0eaa2e9fc96bf472df0247a26bec77bf3ff9
NOTICE Component attributions & license notices 2.4 KB (2,396 B) 06dadecf251ceef34f8579b6088aaa0c959f31e26d3150fd9878abb4ca0d02e5

Download

pip install -U huggingface_hub
hf download litert-community/MediaPipe-DecisionMaker-Laya laya_s256.task --local-dir .
hf download litert-community/MediaPipe-DecisionMaker-Laya laya_s512.task --local-dir .

Inputs, Outputs, and Model Specifications

Bundle Contents (laya_s256.task)

Bundled Member Uncompressed Size SHA-256 (prefix)
model.tflite 250.89 MB (250,889,040 B) 712f2fea6b2c3975…
tokenizer.json 34.36 MB (34,363,188 B) 609d8f4c067cd395…
calibration.json 9.2 KB (9,156 B) 80e148c68154b607…
token_embeddings_fp16.bin 393.22 MB (393,216,000 B) 58608f7bbd72e03a…

Sub-model model.tflite

Direction Tensor Shape Type Notes
Input serving_default_inputs_embeds [1, 256, 768] float32 β€”
Input serving_default_attention_mask [1, 256] float32 β€”
Input serving_default_qtype_onehot [1, 3] float32 β€”
Output serving_default_token_logits_output [1, 256] float32 β€”
Output serving_default_pooled_cls_output [1, 768] float32 β€”

Bundle Contents (laya_s512.task)

Bundled Member Uncompressed Size SHA-256 (prefix)
model.tflite 251.81 MB (251,806,544 B) 2023399e95b4ef15…
tokenizer.json 34.36 MB (34,363,188 B) 609d8f4c067cd395…
calibration.json 9.2 KB (9,156 B) 80e148c68154b607…
token_embeddings_fp16.bin 393.22 MB (393,216,000 B) 58608f7bbd72e03a…

Sub-model model.tflite

Direction Tensor Shape Type Notes
Input serving_default_inputs_embeds [1, 512, 768] float32 β€”
Input serving_default_attention_mask [1, 512] float32 β€”
Input serving_default_qtype_onehot [1, 3] float32 β€”
Output serving_default_token_logits_output [1, 512] float32 β€”
Output serving_default_pooled_cls_output [1, 768] float32 β€”

Use with MediaPipe Tasks

Python Quickstart

pip install -U mediapipe huggingface_hub
import mediapipe as mp
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="litert-community/MediaPipe-DecisionMaker-Laya", filename="laya_s256.task")

BaseOptions = mp.tasks.BaseOptions
DecisionMaker = mp.tasks.decision.DecisionMaker
DecisionMakerOptions = mp.tasks.decision.DecisionMakerOptions
ChoiceQuestion = mp.tasks.decision.ChoiceQuestion

options = DecisionMakerOptions(
    base_options=BaseOptions(model_asset_path=model_path),
    max_num_tokens=256,
)
with DecisionMaker.create_from_options(options) as maker:
    q = ChoiceQuestion(
        criteria={
            "refund": "Requesting a refund or money returned.",
            "tech_help": "Reporting a bug or asking for technical assistance.",
        },
        normalize_prior=True,
    )
    res = maker.evaluate_choice("The same payment was charged twice. Please refund me.", q)
    print("Choice:", res.selected_key, res.probabilities)

Cross-Platform Guides & Samples

Latency Benchmarks

The MediaPipe Decision Maker documentation does not include a standalone latency table. The upstream LiteRT package (litert-community/Laya-Multilingual-LiteRT) reports 51.9 ms per 256-token question on a Galaxy S26 GPU (FP32) and 36.5 ms on the Hexagon v81 NPU (excluding β‰ˆ14.6 ms host embedding lookup).

Intended Use, Limitations, and Fairness

  • Intended applications: Multilingual zero-shot intent routing, policy condition checks, and rubric scoring in a single deterministic forward pass on edge devices (works across both English and non-English / Japanese inputs and complex "unless" rules where bi-encoders struggle).
  • Limitations & license notices: Keep option descriptions concise (5–15 high-signal words) so schema + context fit within the 256- or 512-token window. See the bundled LICENSE and NOTICE files for convaiinnovations/laya (Apache-2.0) and mmBERT-base (MIT) attributions.

Provenance

Every model binary in this repository is byte-identical (unmodified weights, metadata, and .task bundle archive bytes) to the official artifact served by Google from storage.googleapis.com/mediapipe-models.

Repo File Upstream URL GCS Generation Upstream Last-Modified Pinned Bucket Version
laya_s256.task laya_s256.task 1791231584156797 Mon, 05 Oct 2026 20:19:44 GMT /1/
laya_s512.task laya_s512.task 1791232006884043 Mon, 05 Oct 2026 20:26:46 GMT /1/

License and Citation

  • License: apache-2.0 β€” Distributed under Apache License, Version 2.0 (see bundled LICENSE and NOTICE files for component attributions).

Citation

Upstream checkpoint: convaiinnovations/laya (mmBERT-base: jhu-clsp/mmBERT-base); LiteRT conversion provenance: litert-community/Laya-Multilingual-LiteRT.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for litert-community/MediaPipe-DecisionMaker-Laya

Quantized
(95)
this model

Collection including litert-community/MediaPipe-DecisionMaker-Laya