---
language:
  - en
datasets:
  - nvidia/SEED-Timeline-Annotations
base_model:
  - nvidia/ARDY-Core-RP-20FPS-Horizon40
  - sentence-transformers/all-MiniLM-L6-v2
license: other
license_name: ardy-mini-composite-model-terms
license_link: https://huggingface.co/intsuc/Llama-3-ARDY-Mini-Core40-Browser/blob/main/MODEL_TERMS.md
tags:
  - motion-generation
  - text-to-motion
  - onnx
  - onnxruntime-web
  - webgpu
  - browser
  - minilm
---

# Llama 3 ARDY Mini Core40 Browser

**Built with Meta Llama 3.**

Browser-ready ONNX files for ARDY Mini, a browser-ready derivative of NVIDIA
ARDY's Core40 text-to-motion model. Its Llama 3 + LLM2Vec text-conditioning
path is distilled into an all-MiniLM-L6-v2-based encoder that emits the same
1,024-dimensional root and 1,024-dimensional body conditions. The ARDY motion
generator is converted, not distilled, and runs locally with ONNX Runtime
Web's WebGPU execution provider.

- Interactive app: [ARDY Mini](https://huggingface.co/spaces/intsuc/ardy-mini)
- Source and reproducibility code: [intsuc/ardy-mini](https://github.com/intsuc/ardy-mini)
- Publisher/contact: intsuc, [i@intsuc.dev](mailto:i@intsuc.dev)

## Files and browser selection

The app selects one variant automatically. It downloads only the selected
variant and stores verified files in the browser cache.

| Variant | WebGPU requirement | Compressed transfer size |
| --- | --- | ---: |
| `fp16/` | `shader-f16`; mixed FP16 decoder, FP32 denoiser and text encoder | 652.53 MiB |
| `fp32/` | WebGPU; no optional shader feature | 684.30 MiB |

Each directory contains `model.json.gz` plus individually gzip-compressed ONNX
and tokenizer transports declared by that manifest. Raw and compressed byte
sizes and SHA-256 hashes are embedded in the manifests. Top-level
`SHA256SUMS` covers every published file other than itself.

The production app should resolve files from an immutable Hub commit, not
from `main`:

```text
https://huggingface.co/intsuc/Llama-3-ARDY-Mini-Core40-Browser/resolve/<commit>/
```

## Intended use

This release is intended for local, interactive generation and preview of
human character motion from clear English descriptions. It supports the
`ARDY-Core-RP-20FPS-Horizon40` architecture, generates at 20 FPS, and is
designed for the accompanying WebGPU browser runtime.

Reasonable uses include animation prototyping, character previews, games,
creative tools, and motion-generation research consistent with the applicable
licenses and acceptable-use policies.

## Out-of-scope use

This is not intended as a safety-critical physical controller. Do not deploy
it where failure could cause death, bodily injury, or material damage,
including transportation, heavy machinery, critical infrastructure, or
weapons. Simulation and non-safety-critical robotics research are not excluded
by this statement. Uses prohibited by the Meta Llama 3 Acceptable Use Policy,
the NVIDIA Trustworthy AI terms, or applicable law are outside scope.

## Training corpus and transformations

The MiniLM condition encoder was trained and evaluated exclusively on English
motion descriptions from NVIDIA Corporation's
[`nvidia/SEED-Timeline-Annotations`](https://huggingface.co/datasets/nvidia/SEED-Timeline-Annotations),
revision `b2cf916d8ef7a1e49fc4f0ce9e00c1981d3b9d8f`, published under CC BY 4.0.
The pinned `timelines.jsonl` has SHA-256
`379d6a5b86cea06b7201d485d19ee53512cc58449352b3cf113a95d1d27603d8`.

Preparation selects `overview_description` and `events.description`, applies
Unicode NFKC and whitespace normalization, removes descriptions over 512
characters, deduplicates case- and punctuation-insensitively, groups related
recordings, and creates deterministic disjoint train/validation/test groups.
The resulting 64,287 unique descriptions comprise 51,482 train, 6,710
validation, and 6,095 test prompts. Dataset records and derived prompt text are
not distributed in this repository.

## Model lineage

- Motion base: `nvidia/ARDY-Core-RP-20FPS-Horizon40`, pinned in
  `MODEL_PROVENANCE.json`
- Student base: `sentence-transformers/all-MiniLM-L6-v2` at
  `1110a243fdf4706b3f48f1d95db1a4f5529b4d41`
- Teacher foundation: `meta-llama/Meta-Llama-3-8B-Instruct` at
  `8afb486c1db24fe5011ec46dfbe5b5dccdb575c2`
- Teacher MNTP adapter:
  `McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp` at
  `31474e395ada192e8ed1586db6be79fb3b70c9c0`
- Teacher supervised adapter:
  `McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised` at
  `baa8ebf04a1c2500e61288e7dad65e8ae42601a7`

The Llama, LLM2Vec, source MiniLM, original ARDY checkpoint, teacher cache,
and dataset records are not included as separate files. The distributed
graphs contain the converted motion model and trained student parameters
needed by the browser runtime.

## Evaluation

The selected 100-epoch student was chosen on the validation split. On all
6,095 held-out test descriptions, its 2,048-dimensional teacher-condition
cosine was `0.97557` overall (`0.97861` root and `0.97033` body).

Matched motion evaluation uses 64 held-out prompts, three seeds, 80 frames at
20 FPS, identical ARDY weights and noise, and no post-processing:

| MiniLM vs teacher, same seed | Result |
| --- | ---: |
| Global MPJPE | 0.06475 m |
| Root-aligned MPJPE | 0.03140 m |
| Root ADE / FDE | 0.04776 m / 0.08181 m |
| Motion cosine | 0.99635 |
| Foot-contact agreement | 0.96898 |

These metrics measure replacement fidelity to the Llama 3 + LLM2Vec teacher,
not semantic quality against ground-truth motion. They are not comparable to
paper FID or R-precision results.

The adopted mixed-FP16 policy keeps the autoregressive denoiser and text
encoder byte-identical to FP32 and converts only reviewed decoder regions. In
a 64-prompt × 3-seed × 5-window comparison against the FP32 browser graphs,
accumulated 200-frame MPJPE was 0.265 mm mean / 0.350 mm p95, with 99.998%
contact agreement. Full aggregate inputs and environment details are included
under `reports/`.

The FP16 ablation report retains the browser model's internal pre-release ID
and the manifest hashes used during that evaluation. Public-release metadata
changes those manifest bytes but not the evaluated ONNX or tokenizer payloads.
`MODEL_PROVENANCE.json` binds the report to this release using exact raw graph
and tokenizer SHA-256 hashes plus the checkpoint and student identities.

## Limitations

- English motion prompts are the evaluated input domain. Unusual wording,
  misspellings, non-English text, or motion concepts absent from training may
  reduce prompt adherence.
- Generated motion may exhibit foot sliding, contact errors, implausible
  transitions, self-intersections, drift, or prompt mismatch.
- The model is limited to the Core40 representation and the browser runtime's
  fixed generation schedule.
- Results are based on one deterministic training seed and the recorded
  hardware/software environment.
- FP16 and FP32 refer to browser graph precision, not an assurance of identical
  motion in every continuation; see the ablation report for measured bounds.

## Terms

This is a composite model release, not a single Apache-2.0 work. Read
[`MODEL_TERMS.md`](MODEL_TERMS.md), [`NOTICE`](NOTICE), and the complete texts
under [`LICENSES/`](LICENSES/) before use or redistribution.

This model is not affiliated with or endorsed by NVIDIA, Meta, McGill
University, Hugging Face, or sentence-transformers.

## Reproducibility

`MODEL_PROVENANCE.json` records the source commit, exact upstream revisions,
artifact identities, manifest hashes, and report hashes. `SHA256SUMS` provides
a flat integrity index for the published repository. Export and release
commands are documented in the source repository.