QLoRA fine-tune of Qwen/Qwen3.8-27B for generating complete, self-contained Three.js scenes as single HTML files — the task of ccore/Qwen3.5-4B-3D-Coder, trained on ccore/ThreeJS-3D-Scenes (MIT). Part of an architecture-comparison post-training portfolio (same data, same evaluator across LFM3, Qwen3.8, and planned MoE/RWKV tracks).
This repo contains both the merged bf16 model (model-*.safetensors) and the LoRA
adapter it was merged from (adapter_model.safetensors, r=64).
Same protocol for every model in the portfolio: 60 stratified validation prompts
(6 per difficulty level, seed 42), temperature 0.6 / top-p 0.95 / top-k 20,
max 16,384 new tokens. Metrics: complete HTML document returned; pinned Three.js
0.160.0 module import present; extracted module script passes node --check.
| Model | Complete HTML | Pinned import | JS syntax |
|---|---|---|---|
| Gold (dataset targets) | 60/60 | 59/60 | 59/60 |
| Qwen3.8-27B QLoRA (this model) | 59/60 | 59/60 | 58/60 |
| LFM3-1.7B full SFT | 16/60 | 16/60 | 3/60 |
| LFM3-1.7B-Base (zero-shot) | 40/60 | 7/60 | 16/60 |
Failure analysis of this model's outputs: 59/60 generations finish with a complete
HTML document, and 0 of the 59 checkable scripts fail with a duplicate-declaration
error — the failure mode that dominates the 1.7B model's fine-tune (11 of its 13
finished-but-invalid outputs) does not appear at this scale. Raw generations and
scores are in eval/.
<|im_end|>.fde-3dcoder-portfolio).Research/portfolio artifact for Three.js scene generation in the dataset's format. Outputs are single-file HTML scenes; generated code should be reviewed before use in production. The model inherits the base model's general capabilities and limits outside this task; it was not evaluated on general benchmarks.
Apache-2.0, following the base model. Dataset: MIT (ccore/ThreeJS-3D-Scenes).
model-00001-of-00002.safetensors 50 GBmodel-00002-of-00002.safetensors 4.9 GBadapter_model.safetensors 1.9 GBtraining_args.bin 6.2 kB