reaperdoesntknow/Symiotic-14B
SymbioticLM-14B by reaperdoesntknow is a 17.8 billion parameter hybrid symbolic–transformer model built on Qwen-14B. It integrates neural representation with structured symbolic cognition, featuring persistent memory, multi-stage symbolic routing, and self-organizing knowledge structures. This model is optimized for advanced cognitive reasoning, symbolic math, code generation, and scientific simulations in complex problem domains.
Loading preview...
SymbioticLM-14B: Hybrid Symbolic–Transformer Model
SymbioticLM-14B is a 17.8 billion parameter hybrid model developed by reaperdoesntknow, combining a Qwen-14B transformer backbone with advanced symbolic cognition modules. It is designed to excel in symbolic domains by tightly coupling neural representation with structured symbolic processing, aiming to match or exceed the performance of top-tier LLMs in these areas.
Key Capabilities & Architectural Highlights
- Hybrid Architecture: Integrates a Qwen-14B transformer with specialized symbolic modules like ThoughtDynamicsLNN, LiquidThoughtProcessor, CrystallineProcessor (DNAConv GNN), and HelicalDNAProcessor.
- Persistent Memory: Features 4096 symbolic states in FP32, retrieved using entropy and contextual similarity, enabling true memory for multi-step interactions.
- Advanced Reasoning: Supports multi-stage symbolic routing and self-organizing knowledge structures, underpinned by Discrepancy Calculus for dynamic completeness and stability in memory consolidation.
- Dream Mode: Includes a background symbolic simulation for open-ended cognitive processes.
- Intent-Based Routing: Utilizes an intent classifier and entropy gating to select appropriate processing paths.
Intended Use Cases
- Advanced Reasoning Agents: Ideal for complex, multi-step conversational agents requiring true memory.
- Symbolic Generation: Excels in long-form symbolic theorem generation, proof planning, and symbolic math/code synthesis.
- Scientific Applications: Suitable for scientific dialogue, symbolic simulations, and reasoning in fuzzy or discontinuous problem domains.
Limitations
- Memory utility is enhanced with curation and seeding.
- Symbolic cognition is not instruction-tuned for general question answering.
- Increased VRAM usage due to FlashAttention and symbolic modules during generation.