Views
No views yet
A domain-specific AI architecture that generates truly creative, novel names for brands, YouTube channels, social media handles, and more — using Uniform Discrete Language Diffusion instead of autoregressive LLMs.
1# Clone and setup
2!git clone https://huggingface.co/krystv/neurolex-v4-creative-name-diffusion
3%cd neurolex-v4-creative-name-diffusion
4!python setup.py # ← IMPORTANT: fixes imports
5
6# Train (~25 minutes on free T4)
7!python train.py --size base --epochs 30 --batch_size 256
8
9# Generate names
10from neurolex_v4_model import *
11checkpoint = torch.load('./checkpoints/neurolex_v4_best.pt')
12config = NeuroLexConfig(**checkpoint['config'])
13model = NeuroLexV4(config).cuda()
14model.load_state_dict(checkpoint['state_dict'])
15model.eval()
16
17names = model.generate(
18 domain_id=DOMAIN_TO_ID['tech'],
19 style_id=STYLE_TO_ID['sharp'],
20 lang_id=LANG_TO_ID['english'],
21 target_length=8, batch_size=20,
22 cfg_scale=2.5, temperature=0.9,
23 n_steps=80, odd_alpha=8.0, device='cuda'
24)
25print(names)| Problem | Root Cause | Example |
|---|---|---|
| Repetition | AR probability feedback loops | Generates "Nexaflow" 50 times |
| Generic outputs | MLE training → common patterns | "TechFlow", "DataStream", "CloudSync" |
| Mode collapse | Small model memorizes modes | Only 47% uniqueness (v3) |
| Can't invent words | Subword tokenizers recombine known pieces | Just concatenation of morphemes |
| Sounds cringe | No phonotactic awareness | "Xyzptlk", "Blorpify" |
| No cultural sense | Ignores language-specific sound patterns | Same output for Japanese vs French vibe |
LLM/GPT approach (BROKEN):
[Start] → P(next|left) → P(next|left) → ... → same output every time
NeuroLex v4 (WORKS):
[Random Noise] ← denoise ← denoise ← ... ← [Novel Name]
(different noise each time = different output each time)| Innovation | What It Does | Based On |
|---|---|---|
| UDLM | Uniform noise → iterative denoising | MDLM (NeurIPS 2024) |
| Classifier-Free Guidance | Control generation without mode collapse | Discrete CFG (2024) |
| ODD | Batch samples actively repel each other | ODD (2025) |
| adaLN | Condition modulates every layer | DiT (2023) |
| Cosine schedule | More refinement time at low noise | DDPM/MDLM |
| Character vocab | Generate truly novel sequences | ByT5 principles |
| Property | Value |
|---|---|
| Parameters | ~12M (base) |
| Vocabulary | 72 characters (a-z, A-Z, 0-9, specials) |
| Max name length | 24 characters |
| Languages | 25 |
| Domains | 20 |
| Styles | 10 |
| Training time | ~25 min on free Colab T4 |
| GPU memory | <8 GB |
| Target diversity | 90%+ uniqueness |
├── neurolex_v4_model.py # Core UDLM architecture (DiT + CFG + ODD)
├── neurolex_v4_dataset.py # Built-in dataset (25 languages, 20 domains)
├── train.py # Training script (CLI)
├── generate.py # Interactive generation script
├── test_model.py # Validation tests
├── setup.py # Run first to fix imports
├── NeuroLex_v4_Training.ipynb # Complete Colab notebook
└── README.md # This file1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model_id = "krystv/neurolex-v4-creative-name-diffusion"
4tokenizer = AutoTokenizer.from_pretrained(model_id)
5model = AutoModelForCausalLM.from_pretrained(model_id)AutoModelForCausalLM with the appropriate AutoModel class.