bonsai-worlds-8b-lora

Bonsai Worlds 8B level-generation LoRA

A rank-32 LoRA trained to produce compact platform-game level plans, full tile grids, and powerups with constrained abilities and modifiers. Use compact plans for level generation: direct full-grid generation did not pass evaluation. This is an adapter file; it requires the original Bonsai Q1_0 base and the compatible PrismML runtime.

Use

Load bonsai-forge-lora-f16.gguf at scale 1.0 alongside the unchanged prism-ml/Bonsai-8B-gguf base Bonsai-8B-Q1_0.gguf, revision 48516770dd04643643e9f9019a2a349cf26c5dbd, SHA256 284a335aa3fb2ced3b1b01fcb40b08aa783e3b70832767f0dd2e3fdfa134bd54. The tested runtime is PrismML-Eng/llama.cpp revision 5d80cff0b8cb9f2bf823cfc4e71e3abb97f290d6. Use the exact prompts and JSON grammars in evidence/native-prompts.json. The Android sampler applies grammar first, then top-k 20, top-p 0.9, temperature 0.6, and seeded sampling. Compact plans expand through the game's deterministic geometry generator. The app validates output, applies the requested seed, and retains a labeled procedural fallback for rejected or delayed generations.

The original base remains about 1.16 GB. The separate adapter is 174,622,912 bytes. Runtime memory also includes both files, the context cache, and working buffers.

Training

Run bonsai-20260917T222435Z used Unsloth on one Modal H100, the official prism-ml/Bonsai-8B-unpacked checkpoint at 376f381570d6115bc03f82adcfa4af0c7672ae54, frozen BF16 base storage/compute, FP32 trainable LoRA values, rank 32, alpha 64, attention and MLP projections, context 2048, response-only loss, learning rate 0.0002, batch 4 and gradient accumulation 4.

The audited synthetic corpus has 250,000 examples: 125,000 plans, 75,000 full levels/chunks, and 50,000 powerups; 184,156 training rows and 65,844 held-out rows. The run completed 1,000 optimizer steps / 16,000 example presentations, less than one corpus epoch. Semantic combinations and wording families are held out in addition to disjoint seeds. Every dataset chat fits 2048 tokens without truncation. Dataset manifests and hashes are in evidence/.

All 399 tensors / 8,188,548,096 original FP16 checkpoint values survived a FP16→BF16→FP16 storage round trip exactly. This establishes storage-value representability, not identical arithmetic. The exported F16 LoRA is applied separately to the original Q1 base; it is never merged and re-quantized into Q1.

Measured acceptance

Held-out loss (128 examples): 2.185389 before, 0.179760 after. The unconstrained GPU sample has 32 plans, 4 grids, and 12 powerups; the paired comparison uses the same 6 original baseline requests. Full metrics, failures, and seed-copy results remain in evidence. Unconstrained GPU output was 32/32 valid and 32/32 prompt-compliant plans, and 12/12 valid and 12/12 prompt-compliant powerups. Direct full grids were only 1/4 valid and 0/4 prompt compliant. Full-grid output is not accepted for gameplay; Android generates levels by expanding compact plans.

Native Q1-plus-LoRA results: 16/16 valid plans, 16/16 prompt-compliant plans, 12/12 valid powerups, and 12/12 prompt-compliant powerups. Actual JavaScript expansion plus independent Python physics checks passed 16/16 plans across 64 finite and infinite contexts, with 0 geometry errors. Android emulator import, native plan/powerup inference, and game UI acceptance passed using this exact adapter checksum.

These are small held-out samples, not a guarantee for arbitrary prompts. Static route validation does not prove dynamic enemy/boss timing. Emulator and CPU-worker timings are not physical Android-phone performance measurements. This model generates structured recipes within the supplied game mechanics, not executable code or unrestricted new mechanics. No proprietary game assets are included.

Adapter SHA256: 85cf09c4c6ddfde3c9e83a69de338a7f17a4c70e93d5a4bb83b65f9589897d63. The manifest identifies all public files and acceptance evidence. Base model and adapter are provided under Apache-2.0.