Drop Files into Corpus (Any Format)
1-D Bytes · 2-D 4×4 Patches · μ-law · AuraGraph v0.4 Tiles
Drop files here or click to browse — any format trains as bytes or 2-D patches
TXT · HTML · JSON · SVG · JS · C · BMP (4×4 Patches) · WAV (μ-law or AuraGraph Tile+Helix) · .AURA.PNG · ANY EXTENSION
WAV Path
BMP Path
Built-In Seeded Test Corpora (Tier 4 Proof & Baselines)
80/10/10 Contiguous Train / Val / Test Split
Corpus Seed
Count / Fmt
Masked Byte Diffusion Training (MDLM Continuous-Time NELBO)
0.0 steps/s
Model Preset
Seed
Schedule
Kernel
Peak LR
v0.4 Recipe
Ready. Load files or click "+ All Seeded Corpora" in Corpus tab, then press START.
Step NELBO (■ train · ■ rolling avg)
Train: — bpb · Held-Out Val: — bpb
Held-Out Likelihood vs Byte N-Gram Baselines (Tier 1 & Tier 4)
Click Measure Held-Out BPB to fit Unigram, Bigram, and Order-4 Kneser-Ney baselines on the train split and compare against Masked Byte Diffusion NELBO on the held-out split.
Generation & Infilling Parameters
Window = 512 bytes (128B overlap context) · 2-D 4×4 Patches
Format
Tags
Gen Seed
Length
bytes (windows of 512 with 128-byte fixed context overlap)
Unmasking
steps
Refinement
Infilling
—
Output Preview
No output yet. Train the model and press Generate.
Audio ready.
Compiler-as-Validator & Repair Report (Tier 3 & Tier 4)
Generate a file to run its native parser, structural checks, and LCS memorisation analysis.
No output
Validity Rate by Format (Raw vs Repaired)
No generations recorded yet. Generate samples or click "Run 10-Sample Validity Benchmark" in Generate tab.
Model Statistics & Likelihood Bound
Total Steps
0
Parameters
—
Corpus Windows
0
Best Train BPB
—
Held-Out Val BPB
—
Window Size
512B / 64P
Vocabulary
256 + [MASK]
Active Preset
Laptop
Architecture v0.4: Learned Bidirectional Masked-Byte Transformer
Forward Process: At diffusion time
t ∈ (0, 1], each byte x₀[i] ∈ {0..255} is independently replaced by [MASK] (256) with probability μ(t) = 1 − α(t).Continuous-Time NELBO Bound (bits per byte):
bpb(x₀) = E_{t~U(0,1), x_t} [ (1 / (L ln 2)) · (−α'(t)/(1−α(t))) · Σ_{i : x_t[i]=[MASK]} −log p_θ(x₀[i] | x_t, t, c) ]
Bidirectional Transformer Denoiser: Token + 1D/2D (row, col) + 4×4 sub-patch slot embeddings → N× [Pre-AdaLN → RoPE Multi-Head Bidirectional Self-Attention + Local Conv → Pre-AdaLN → SiLU FFN] → Final LN → Differentiably-Gated Bilateral Context + 256-way Byte Head + 4-way Timing Helix Event Head.
Presets (v0.4): Phone (2L, 64d, 4H, 676,325 params) · Laptop (3L, 128d, 4H, 1,459,237 params) · Patient (4L, 192d, 6H, 2,818,021 params) with Resident WASM/SIMD
matmul_f32, Fast Typed-Array BilateralContextMemory, Out-of-Fold Memory Training, and Per-Position Hidden Gate.Weight Files (.fabula v0.4)
⚠ v0.4 uses Masked Byte Diffusion (FabulaTransformer v0.4). Legacy v0.1 and v0.2 Gaussian MLP .fabula files are detected and rejected gracefully with a clear diagnostic message.
Live Console
Gradient Health
Per-parameter RMS grad norm · ■ healthy · red=vanishing · yellow=spike
—
Format Step Distribution
Optimizer & Kernel State
Generation Journal (newest first · max 50)