About the Project
Osakra Research is an independent, solo research effort exploring continuous-time dynamical systems, latent recurrence, and vector-symbolic architectures in compact language models[cite: 6, 8].
Rather than training frontier-scale foundation models on massive clusters, this project investigates targeted architectural questions under strict local hardware constraints: can compact open backbones deliberate directly within continuous hidden vector space ($\mathbf{h} \in \mathbb{R}^d$) on a single consumer laptop GPU (< 4.5 GB VRAM)[cite: 6, 8]?
Research Scope & Limitations
- Continuous Latent Deliberation: Investigating whether variable-depth stochastic recurrence ($k \in [2, 16]$) and attractor regularization ($\mathcal{L}_{\text{drift}}$) can stabilize continuous latent Chain-of-Thought reasoning without emitting verbose textual tokens[cite: 6, 8].
- Dynamical Systems Modulation: Experimenting with continuous-time Liquid Time-Constant (LTC) neural ODEs and Holographic Reduced Representations (HRR) for associative memory binding[cite: 5, 6, 8].
- Grounded Scope: This is exploratory, single-author research[cite: 6, 8]. These models do not compete with 70B+ flagship models on broad open-domain knowledge[cite: 6, 8]. The goal is to document empirical findings, trade-offs, and failure modes when compute is spent in latent space rather than token space[cite: 6, 8].
Releases & Verification
- Project Norn V15: An exploratory 4.45B parameter neuro-symbolic checkpoint testing continuous latent Chain-of-Thought deliberation on top of a 4-bit quantized base[cite: 6, 8].
- Evaluation Suite: Bundles all 250 held-out evaluation samples and scoring harnesses directly within the repository to allow full local reproducibility[cite: 6, 8].
All code, adapters, and data are published under the Apache-2.0 license to encourage open experimentation and peer verification[cite: 6, 8].