π Taylor-Data: Large-Scale Neural Dynamics & Optimization Dataset (V4)
Taylor-Data is a comprehensive, large-scale dataset containing 2 Billion parameter transitions and loss landscape interactions collected across diverse neural network architectures, optimization regimes, and multi-scale horizons.
It is designed for training World Models of Neural Dynamics, Meta-Optimizers, Quasi-Newton Credit Assignment Networks, and Continual Learning Controllers.
π Overview & Purpose
Traditional optimization operates locally via first-order derivatives without holistic awareness of loss landscape geometry, higher-order curvature, or multi-step trajectory evolution.
Taylor-Data provides high-fidelity, structured trajectory transitions to train models capable of:
- Modeling second-order loss curvature and Taylor series consequences $\Delta L(\theta, \delta)$.
- Predicting analytical derivative dynamics and chain-rule transitions without explicit backpropagation.
- Simulating multi-horizon kinematic parameter evolution across complex loss landscapes.
- Performing selective, surgical parameter interventions and continual learning.
π¦ Dataset Composition (2 Billion Transitions)
The dataset is partitioned into three specialized subsets:
| Subset Name | Volume | Description |
|---|---|---|
base_moe_v4 |
1,000,000,000 | Canonical 148-D state representations paired with multi-directional geometric probes (Newton, Hessian, Orthogonal, SPSA, NES, LoRA, Langevin) and true $\Delta L$ consequences. |
expert_e8_derivatives |
500,000,000 | Analytical chain-rule derivative transitions and gradient vector mappings under parameter displacement steps across real architectures. |
expert_e9_forward |
500,000,000 | Multi-horizon kinematic rollouts ($H \in {4, 8, 16, 32, 64, 128}$) with adaptive Brownian diffusion and noise scale dynamics $\sigma \propto \sqrt{H}$. |
π¬ Feature Schema
Each sample is stored in high-performance Parquet format (Zstandard compression) containing 32-dimensional canonical coordinate blocks:
s_148/s_0_148(FixedSizeList[float16, 148]): Compact 148-D kinematic state vector (normalized weights, velocities, accelerations, momentum, spectral rank, SNR, Rayleigh quotient, loss EMA history, and layer topology metadata).delta_32/dW_32/dW_accum_32(FixedSizeList[float16, 32]): Parameter perturbation and action vectors.delta_L(float32): Ground-truth scalar loss variation observed on the target landscape.grad_32(FixedSizeList[float16, 32]): True analytical gradient vectors for derivative modeling.s_target_148(FixedSizeList[float16, 148]): Future state vector reached after horizon $H$.horizon_H(int16): Step horizon ($4 \le H \le 128$).noise_level(float32): Applied perturbation magnitude during the rollout trajectory.
βοΈ Optimization & Landscape Diversity
The transitions encompass a wide spectrum of synthetic and real optimization dynamics:
- Landscape Geometries: Ill-Conditioned QR Haar ($\kappa \in [1, 10^6]$), Rosenbrock ravines, Barren Plateaus, High-Order Polynomial Wells, Non-Smooth $L_1$/Cusp frontiers, and Discrete Quantized landscapes.
- Optimization Algorithms: Momentum SGD, Nesterov, AdamW, Muon (Newton-Schulz Matrix Orthogonalization), SPSA (Zero-Order), NES (Evolutionary Strategies), and Langevin Diffusion.
π License & Terms
This dataset is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC-BY-NC-ND-4.0).
- Downloads last month
- 185