R4T DIFFUSION
Continuous Diffusion Retriever • Experiment Journal
Total Sweeps & Runs
-
SQLite WAL Multi-Process Journal
Best Validation Loss
-
EDM Weighted MSE
Peak Pairwise Diversity
-
Cosine Distance across 10 Slots
Sampling Latency (RTX 4090)
-
16-Step Heun ODE / 10 Vectors
W&B Sweep View

Multi-Parameter Parallel Coordinates (Rank Chart)

Drag vertical axis pills left or right to reorder pillars. Drag vertically along any axis to brush-filter runs. Hover over splines to inspect configuration & metrics. Click any line to jump to decoded subqueries.

💡 Tip: Drag an axis title to reorder pillars. Brush along any axis to filter runs. Active: 8 / 8 runs. Lines colored by Rank (Best → Low)
Hyperparameter Sweep Ranked Leaderboard
8 runs ranked
Rank Run Name Depth Learning Rate Slot Ordering Best Val Loss Prompt Alignment Diversity Latency Actions
Training Convergence Trajectories (All Runs)
Val Loss vs Epoch
Multi-Directional Fanout Pareto Frontier (Rank Chart)
Alignment vs Diversity
Validation Loss vs Single-Pass Sampling Latency
Quality / Speed Trade-off
Diffusion Sampling Velocity Spectrum (RTX 4090)
Throughput (Sets/sec)
Continuous Diffusion Retriever Architecture Milestones WAL Journal Mode Verified
Milestone 1 • Small Corpus Baseline
14M Backbone (54k Items)
Best Val Loss: 0.4922 (Epoch 51)
Prompt Alignment: 0.473 Cosine
Sampling Latency: 264 ms (RTX 4090)
Status: Overfitting Small 538-Category Set
Milestone 2 • Capacity Scaling
54M Backbone (12 Layers)
Best Val Loss: 0.4893 (Epoch 49)
Prompt Alignment: 0.482 Cosine
Sampling Latency: 540 ms (RTX 4090)
Status: Higher Latency, Marginal Delta
Milestone 3 • Full Corpus Drift
Unordered Slots (540k Items)
Epoch 2 Val Loss: 0.6929 [Peak]
Epoch 50 Val Loss: 0.8558 [Drift]
Root Cause: Slot Permutation Blur across 10 Slots
Status: Target for Canonical Sorting Sweep
Milestone 4 • Parameter Sweep Active
Canonical Cosine-Sorted Target Set
Hypothesis: Eliminates Post-Epoch-2 Drift
Architecture Grid: 2-Layer / 4-Layer / 6-Layer
Learning Rate Grid: 1e-4 / 3e-4 / 5e-4
Visualizer: WandB Output Inspector Active
Run Name ⇅ Experiment ⇅ Status ⇅ Layers ⇅ Slot Ordering ⇅ Best Val Loss ⇅ Alignment ⇅ Diversity ⇅ Latency (ms) ⇅ Started ▼ Actions

Select a Run

Completed Diffusion
Experiment Group

W&B Model Output Inspector • Decoded Multi-Directional Fan-Out Queries

Hardware Performance & Profiling Notes
No hardware benchmark recorded for this run.
Run Configuration JSON
{}
1-Step Consistency Distillation 1.64 MB INT4 Model 169 µs Core Latency

Interactive Live Query Fanout Engine

Execute instant single-step latent continuous diffusion fanouts directly on RTX 4090 hardware Tensor Cores.

Type a query above or click one of the quick suggestions to generate instant 10-slot fanouts!