PeytonT/SimulateAnything-Gemma-3-1B-v1

SimulateAnything Gemma 3 1B

A real rank-16 PEFT LoRA for bounded scene authoring and structured request interpretation. This is a derivative of Google Gemma 3 1B IT, with 13,045,760 trainable adapter parameters. The merged q4 ONNX model runs locally in a WebGPU browser. The separate adapter is in adapter/.

Status: experimental browser-qualified scene authoring. SimulateAnything pins tested artifact revision 3656f70ba19d67ed278b0ffc9db73d64e282c2a9; later commits update evidence only. The graph, external weights and adapter hashes were verified against published LFS identities. Training success and native outputs alone were not used as the deployment gate.

Actual browser q4 checks	Baseline Gemma 3 1B	This fine-tune
Fixed 35-check panel	8/35	34/35
Prior 10-case retention panel	1/10	6/10

These are finite mechanical checks, not a 97% general correctness claim. The main panel includes 28 authoring/conversation checks and seven application constraint checks. The candidate scored 27/28 and 7/7 respectively; the baseline scored 2/28 and 6/7, rejecting even the correct draft. The first-generation negated prompt still produced an action, but separate request extraction returned no action; the actual multi-turn UI preserved the saved scene. The original four-purple-spheres/5m request passed actual browser generation, explicit Apply and native free-fall verification. No chat text was observed in outgoing requests.

Known weaknesses: exact XYZ/custom-hex-color copying, some retained-property edits, centimeter conversion, and consistent unsupported-request refusals. Unsupported tree/code shapes were rejected by the compiler. Learned extraction can agree with a wrong draft. These results do not establish the best general model or arbitrary simulation reliability. Full outputs, rejected candidates and limitations are retained in qualification.json and the source repository.

The model proposes declarative rigid-body scenes, visual forest/creek scenery, and edits. A second request-only inference extracts object counts and constraints; application code compares these with the compiled scene. SimulateAnything uses its existing AgentCore/EnvironmentActions lifecycle and native Rapier dry-run before explicit user review and Apply. Generated code is never executed. Learned extraction can still be wrong; this is not a proof of user intent or physical accuracy.

Training and scope

First pass: 1,232 original synthetic authoring/conversation examples, three epochs. Second pass: 2,464 examples, one epoch, replaying authoring with structured-intent extraction. The exact four-purple-spheres-at-five-meters tuple is excluded from training but is an exposed regression test. Seed 716, rank 16, alpha 32, dropout 0.05, AdamW at 0.0002 with cosine decay. Training loss is not accuracy. See provenance.json for hashes and versions.

Supported scene actions currently cover spheres, boxes, cylinders, mass, pose, size, restitution, rigid contacts and visual scenery. These actions do not solve hydrology, chemistry, ecosystem growth or arbitrary multiphysics. The model is not conscious, does not train online, and cannot guarantee a correct interpretation. This is supervised domain adaptation, not general embodied reinforcement learning.

Reproduction

Source, training-data builders, native rollout checks, export code and qualification scripts are pinned in SimulateAnything. Use the prompts from web/core/chat-actions.js and web/core/scene-author-agent.js, greedy decoding and a 512-token budget. Browser runtime: Transformers.js 4.3.2-derived Gemma-only build with ONNX Runtime Web. Export: Transformers 4.57.6, PEFT 0.17.1, Optimum ONNX 0.1.0/Optimum 2.0.0, ONNX Runtime 1.21.0; symmetric q4 blocks of 32.

The ONNX graph and external data must be distributed together. The adapter can be loaded using PEFT with subfolder="adapter"; pin a model-repository commit and the base revision from provenance.json.

License and modifications

All modified weights are identified in NOTICE, ONNX graph documentation and adapter metadata. Use and distribution are governed by the included LICENSE and NOTICE, including the Gemma use restrictions and prohibited-use policy. No Google endorsement is implied.

The deployed SimulateAnything workspace also passed the actual-model UI suite after a fresh Hugging Face download: the original four-sphere request, review/apply, native motion, conversation/negation preservation, cancellation and no outgoing chat text. Deployment and versioned-source identities are retained in qualification.json.