🧠 Cartridge_084: Semi-Plastic Cognitive Architecture (Rank-16 Micro-LoRA Hypernetwork)

Cartridge_084 is an omni-agnostic, zero-hardcoded, semi-plastic language model system. It combines a frozen 16M parameter foundational backbone with an ultra-lightweight Rank-16 Multi-Layer Micro-LoRA Hypernetwork to dynamically project fast synaptic weights ($\Delta W \in \mathbb{R}^{384 imes 384}$) on-the-fly directly from 32-dimensional semantic vectors.

---

🏛️ Architecture Overview

The system implements a dual-cortex dynamic cognition loop:

• ❄️ Cold Intelligence (Frozen 16M Backbone + 32-d MRL Space): The foundational Transformer (model_v8_4_grpo.pt) remains 100% frozen (requires_grad = False). It preserves perfect syntactic integrity, discourse memory anchoring (ToC), and zero catastrophic forgetting ($\Delta ext{logit} = 0.000000$).
• 🔥 Hot Intelligence (Instantaneous Synaptic Plasticity): Queries are projected into a 32-dimensional Matryoshka Representation Learning (MRL) latent space. The hypernetwork (rank16_hypernet.pt) instantly generates low-rank adapters ($A \in \mathbb{R}^{384 imes 16}, B \in \mathbb{R}^{16 imes 384}$, ~48 KB) that dynamically steer the backbone hidden states ($\Delta h = 2.5 \cdot (h A) B$).

flowchart LR
    Q["User Query"] --> MRL["32-d Static MRL Embedder"]
    MRL -->|z in R^32| HYP["Rank-16 Hypernetwork"]
    HYP -->|A: 384x16, B: 16x384| INJ["Dynamic Fast-Weight Injection"]
    Q --> TOK["Tokenizer (SentencePiece)"]
    TOK --> BB["Frozen 16M Backbone (EnhancedSemanticLM)"]
    INJ -->|Δh = 2.5 * (h A) B| BB
    BB --> OUT["Neural Generated Output"]

---

🚀 Key Specifications

Metric / Parameter	Value
Backbone Parameters	16,384,000 (16M)
Backbone Hidden Dimension ($d$)	384
Dynamic Synaptic Adapter Rank	Rank-16 ($384 imes 16 + 16 imes 384 = 12,288$ floats)
Adapter Storage Footprint	~48.0 KB per cartridge / context
Fast-Weight Generation Latency	< 0.5 ms on GPU
Vector Index Latency	~6 - 15 ms end-to-end
Catastrophic Forgetting	0.000000 (Strictly Frozen Backbone)

---

🧪 Verified Benchmark Performance

Query	Model Output	Dynamic Mode	Latency	Status
capital of brazil	Brasilia	Rank-16 Micro-LoRA	142 ms	✅ Pass
capital of turkey	Ankara	Rank-16 Micro-LoRA	6.4 ms	✅ Pass
who is Trump?	Donald Trump	Rank-16 Micro-LoRA	15.6 ms	✅ Pass
capital of usa	Washington	Rank-16 Micro-LoRA	6.3 ms	✅ Pass
ankara türkiyenin...	başkentidir	Rank-16 Micro-LoRA	9.7 ms	✅ Pass
Türkiye'nin başkenti neresidir?	Ankara	Rank-16 Micro-LoRA	6.3 ms	✅ Pass
Fransa'nın başkenti neresidir?	Paris	Rank-16 Micro-LoRA	6.1 ms	✅ Pass
Su kaç derecede kaynar?	100 derecede kaynar	Rank-16 Micro-LoRA	27.8 ms	✅ Pass

---

💻 Quickstart & Inference

1. Clone & Setup

git clone https://huggingface.co/dcx514ai/Cartridge_084
cd Cartridge_084
pip install torch sentencepiece safetensors tokenizers numpy

2. Interactive Chatbot CLI

python chatbot.py

3. Python API Usage

import os
import torch
from chatbot import DynamicPlasticChatBot

ckpt = "checkpoints/model_v8_4_grpo.pt"
spm_model = "checkpoints/spm/tr_unigram_8192.model"

bot = DynamicPlasticChatBot(ckpt_path=ckpt, spm_path=spm_model, mode="lora", device="cuda")

response, telemetry = bot.step("capital of brazil")
print("Response:", response)
print("Telemetry:", telemetry)
# Output:
# Response: Brasilia
# Telemetry: Yarı-Plastik Rank-16 Micro-LoRA (CosSim: 1.000 | 48.0 KB | 6.5 ms)

---

📜 Citation & License

• License: Apache 2.0
• Author: dcx514ai
• Architecture: EnhancedSemanticLM v8.4 + Rank-16 Micro-LoRA Hypernetwork