arabsya87sy7a8

newmodel

Deploy Dedicated

README

License: apache-2.0

Reference

mrshu/claude-fable-5-bratislava-castle-oneshot

The original is a single-shot Three.js application with:

11 modular JavaScript files + 1 HTML = 1,971 lines total
Custom shaders (sky dome, water)
Procedural canvas-painted textures
Generative 20-second WebAudio score
Cinematic camera tracks with banking
Architectural fidelity: 4-winged palace, Crown Tower, Honorary Courtyard, Svatopluk statue, baroque garden, fortification ring, Sigismund/Vienna gates, Most SNP bridge, St. Martin's Cathedral, Michael's Gate, Old Town Hall, Slavín, parliament, Petržalka

Training Approach

This adapter was trained to memorize the structure of the reference project:

Per-file samples (11 JS files + index.html) as separate training examples
Each sample teaches model one module of the reference architecture
LoRA r=64 (high capacity) for memorization
8 epochs to reinforce patterns

Training Details

Table with columns: Attribute, Value
Attribute	Value
Base model	`Qwen/Qwen2.5-Coder-0.5B-Instruct`
Training data	Reference Bratislava castle (11 JS + index.html) + codellama-threejs + synthetic
Total samples	25
Method	QLoRA (4-bit NF4 + LoRA r=64)
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
LoRA alpha	128
LoRA dropout	0.05

How to Use

python

import torchfrom peft import PeftModelfrom transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig # 4-bit quantization (matches training)bnb_config = BitsAndBytesConfig(    load_in_4bit=True,    bnb_4bit_quant_type="nf4",    bnb_4bit_compute_dtype=torch.bfloat16,    bnb_4bit_use_double_quant=True,) tokenizer = AutoTokenizer.from_pretrained("arabsya87sy7a8/newmodel", trust_remote_code=True)base = AutoModelForCausalLM.from_pretrained(    "arabsya87sy7a8/newmodel",    quantization_config=bnb_config,    device_map="auto",    trust_remote_code=True,)model = PeftModel.from_pretrained(base, "arabsya87sy7a8/newmodel")model.eval() # IMPORTANT: Use system prompt for best resultsmessages = [    {"role": "system", "content": "You are an expert Three.js developer. Output complete, modular code. Use separate files when appropriate (terrain.js, sky.js, castle.js, city.js, cameras.js, cinematic.js, audio.js, textures.js, util.js, main.js, index.html). Include custom shaders, procedural textures, and cinematic camera system."},    {"role": "user", "content": "create the most realistic and accurate 3D model of the Bratislava castle..."},]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.no_grad():    output = model.generate(        **inputs,        max_new_tokens=4000,        do_sample=True,        temperature=0.4,        top_p=0.9,        pad_token_id=tokenizer.pad_token_id,        repetition_penalty=1.05,    )print(tokenizer.decode(output[0], skip_special_tokens=True))

Why Qwen2.5-Coder-0.5B-Instruct (NOT FlowerTune)?

The ethicalabs/FlowerTune-Qwen2.5-Coder-0.5B-Instruct variant is a federated SFT on short Python snippets (CodeFeedback/OSS-Instruct style). This actually degrades the model's ability to produce long, structured HTML/JS artifacts. The base Qwen/Qwen2.5-Coder-0.5B-Instruct (HumanEval 61.6) is the correct starting point.

What to Expect

The model has memorized the Bratislava castle reference structure. With the verbatim prompt from the original Claude Fable 5 session, it should produce output that resembles the reference architecture:

✅ Three.js code (THREE.Scene, PerspectiveCamera, WebGLRenderer)
✅ Modular file structure (when prompted for it)
✅ Castle components (walls, towers, courtyard, garden)
✅ Camera system (orbit + fly mode)
✅ 20-second showcase animation
⚠️ Simpler geometry than original (model size limit)
⚠️ May have syntax errors (test before using in production)
❌ Will NOT reproduce 1971 lines exactly (0.5B model limit)

Limitations

Model size: 0.5B parameters is fundamentally limited vs Claude's 200+B
Memorization vs generation: Model learns patterns but cannot perfectly reproduce 1971 lines from a single fine-tune
Output length: Even with max_new_tokens=4000, output may be truncated
Reference is 82K chars: Cannot fit in single training sample; trained per-file (11 samples) instead
No shaders/textures/audio: Model may attempt but quality will be lower

Comparison with Original Reference

Table with columns: Aspect, Claude Fable 5 (original), This model
Aspect	Claude Fable 5 (original)	This model
Lines of code	1,971	~500-1500 (truncated)
Modular files	11 JS + 1 HTML	1-3 files typically
Custom shaders	Yes (sky, water)	Maybe (basic)
Procedural textures	Yes (canvas-painted)	Unlikely
Generative audio	Yes (WebAudio)	Unlikely
Cinematic camera

Training Pipeline

Base: Qwen/Qwen2.5-Coder-0.5B-Instruct
Data: 11 per-file samples from reference Bratislava castle (1971 lines total) + 12 codellama-threejs + 1 synthetic
SFT with QLoRA (4-bit NF4 + LoRA r=64, alpha=128, 7 target modules)
8 epochs for memorization
max_seq_length=4096 to fit individual file samples
Auto-upload to HF Hub on completion

Citation

bibtex

@misc{mrshu_bratislava,  title={Bratislava Castle — a cinematic 3D model},  author={Marek Šuppa (mrshu)},  howpublished={https://github.com/mrshu/claude-fable-5-bratislava-castle-oneshot},  note={Originally created by Claude Fable 5}} @misc{qwen25coder,  title={Qwen2.5-Coder Technical Report},  author={Qwen Team},  howpublished={https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct},}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

arabsya87sy7a8

Model Tree

Base

Qwen/Qwen2.5-Coder-0.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities