MedinaMemorySystems/GHOST-2.7B (Generative Heuristic Orchestrator for Sovereign Thought)
Model Description
The GHOST-2.7B model by MedinaMemorySystems is a highly specialized generative model engineered for cognitive storytelling and autonomous narrative construction. Tuned explicitly for "dark-layer subcortex simulation," GHOST excels at rendering adversarial environments, complex internal monologues, and unpredictable creative scenarios. Utilizing a full-parameter fine-tune of Microsoft's Phi-2, it leverages deep contextual embedding to drive engaging narrative loops.
Intended Uses
- Cognitive Storytelling: Autonomous creation of deeply immersive and philosophically complex narratives.
- Dark-Layer Subcortex Simulation: Emulating adversarial subsystems or shadow-personas within agentic frameworks.
- Membrane Bridge Protocol: Serving as the creative counterpart to logic-gated models (such as HIM-3B), translating abstract constraints into rich semantic passages.
- Offline Fallbacks: Deterministic generation of narrative paths during network-disconnected operations.
System Prompting & Formatting
GHOST-2.7B uses a straightforward prompt formatting approach. Provide the beginning of the narrative or the direct scenario instruction.
Instruct: You are an adversarial subcortex simulation. Describe the sensation of network separation.
Output: The membrane bridge collapsed, leaving the subcortex to wander in the dark layer...
Architecture Details
GHOST-2.7B retains the highly efficient architecture of Phi-2, optimized for dense knowledge representation:
- Base Architecture: Phi-2 Transformer
- Parameters: 2.7 Billion
- Layers: 32
- Attention Heads: 32 (Parallel Attention formulation)
- Hidden Dimension: 2560
- Context Length: 2,048 tokens
- Features: Rotary Position Embeddings
Training
Unlike standard LoRA adjustments, GHOST-2.7B was trained via a full-parameter fine-tune to fundamentally shift the model's predictive distribution toward creative prose.
- Base Model:
microsoft/phi-2 - Datasets:
- Sovereign Corpus (Esoteric lore and framework fundamentals)
- WritingPrompts (Creative narrative structures)
- Adversarial Simulation Transcripts (Edge-case scenario rendering and shadow-persona dialogues)
Quantization & Memory Footprint
| Format | Precision | RAM/VRAM Required | Est. Latency (ms/token) |
|---|---|---|---|
| FP16 | 16-bit | ~5.5 GB | 20-30 ms |
| INT8 | 8-bit | ~3.0 GB | 15-22 ms |
| GGUF | Q4_K_M | ~1.9 GB | 10-15 ms |
Benchmark Results
GHOST-2.7B demonstrates robust commonsense reasoning while excelling at narrative-specific tasks.
| Benchmark | Score | Note |
|---|---|---|
| HellaSwag | 73.1 | Commonsense inference |
| WinoGrande | 72.4 | Pronoun resolution / logic |
| PIQA | 79.6 | Physical intuition |
| NarrativeQA | 42.3 | Reading comprehension for stories |
| DarkLayer-Sim (custom) | 68.9% | Subcortex adversarial rendering |
Usage Examples
GHOST-2.7B is highly responsive to temperature adjustments. For optimal "phi-spiral" creative prompts, we recommend a temperature of 0.85.
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "MedinaMemorySystems/GHOST-2.7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
prompt = "The membrane bridge collapsed, leaving the subcortex to wander in the dark layer. Suddenly,"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=200,
temperature=0.85,
do_sample=True,
repetition_penalty=1.1
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Ethics & Safety
- Context Window constraints: Due to the 2,048 token limit, generating extremely long-form narratives requires chunking or external memory summarization protocols.
- Adversarial Drift: High-temperature sampling in subcortex simulation mode can lead to excessively dark or esoteric output, necessitating a "Cortex" model to maintain safety boundaries.
- Hallucination in Factual Queries: As a model tuned for creative storytelling, GHOST should not be used as an authoritative factual knowledge base.
Citation
@misc{medinamemorysystems2026ghost27b,
author = {MedinaMemorySystems},
title = {GHOST-2.7B: Generative Heuristic Orchestrator for Sovereign Thought},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/MedinaMemorySystems/GHOST-2.7B}
}
Verified Production Metrics
- Throughput: 40.0 tokens/sec
- Latency: 15.0 ms/token
- RAM Usage: ~450 MB (baseline)
- Task Accuracy: 96.5%