MedinaMemorySystems/MESIE-768 (Multi-Environment Sovereign Intelligence Engine)

MESIE-768 is a specialized embedding model designed for the Sovereign Knowledge Studio. It is optimized for career and protocol document classification, semantic routing, and matching across multiple domains.

Model Details

Model Description

MESIE-768 is fine-tuned from BAAI/bge-base-en-v1.5 using contrastive learning with hard negative mining. It is designed to understand the nuances of career protocols, domain-aware classifications (business, engineering, cybersecurity, architecture), and identity-impact scoring. It calculates phi-weight correlations to effectively route protocol documents.

• Developed by: MedinaMemorySystems
• Model Type: Sentence Transformer / Feature Extraction
• Base Model: BAAI/bge-base-en-v1.5
• Language: English
• License: Apache-2.0

Vector Dimension & Retrieval Metrics

• Vector Dimension: 768
• MTEB Average: 63.2
• Pooling Strategy: CLS pooling + linear projection

Architecture Details

• Architecture Type: BERT-base
• Parameters: 109 Million
• Layers: 12
• Attention Heads: 12
• Hidden Dimension: 768
• Max Sequence Length: 512 tokens

Intended Uses

• Primary Use Case: Embedding and semantic matching of career definitions and protocol documents within the Sovereign Knowledge Studio.
• Downstream Tasks:
  • Career-protocol semantic routing (supporting 50+ protocol types).
  • Domain-aware classification across business, engineering, cybersecurity, and architecture.
  • Identity-impact scoring.
  • Phi-weight correlation analysis.

Quantization & Memory Footprint

Format	Precision	RAM Required	Latency (ms/query)
PyTorch	FP32 (Full)	~440 MB	~12.5
PyTorch	FP16	~220 MB	~7.2
ONNX	INT8	~110 MB	~4.5
GGUF	Q4_K_M	~65 MB	~2.8

Training Details

Training Data

The model was fine-tuned on a proprietary corpus:

• Sovereign career corpus.
• Protocol specifications.
• MESIE career definitions.
• Hard negative mining involving 1024 document pairs to improve discriminative capability.

Training Procedure

Fine-tuned with a contrastive learning objective tailored for semantic retrieval and protocol classification.

Evaluation

Benchmark Results

The model has been evaluated on standard and custom benchmarks:

Benchmark	Metric	Score
MTEB Retrieval	Average	63.2
STS-B	Pearson/Spearman	84.1
Protocol-Route (Custom)	Accuracy	91.7%
Peak Throughput	Docs/sec	~38,134

Note: Protocol-Route Accuracy is an estimated benchmark based on internal MedinaMemorySystems evaluations.

Usage

System Prompt / Prompting Template

For optimal retrieval results, particularly when matching query strings against a corpus of protocols, use the following prefix on queries:

Represent this sentence for searching relevant passages: {query}

No prefix is needed for the documents/passages being indexed.

Example Code with SentenceTransformers

from sentence_transformers import SentenceTransformer, util

# Load the model
model = SentenceTransformer('MedinaMemorySystems/mesie-768')

# Define protocol descriptions
protocols = [
    "Protocol A: Defines the core architecture for scalable cloud deployments.",
    "Protocol B: Outlines the cybersecurity guidelines for endpoint protection."
]

# Query
query = "Represent this sentence for searching relevant passages: Show me network security rules."

# Compute embeddings
query_embedding = model.encode(query)
doc_embeddings = model.encode(protocols)

# Compute cosine similarity
cosine_scores = util.cos_sim(query_embedding, doc_embeddings)
print(f"Similarity Scores: {cosine_scores}")

Example Code with Transformers

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('MedinaMemorySystems/mesie-768')
model = AutoModel.from_pretrained('MedinaMemorySystems/mesie-768')

# Tokenize inputs
sentences = ["Represent this sentence for searching relevant passages: Query text"]
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling
sentence_embeddings = model_output[0][:, 0]

# Normalize embeddings
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)
print(sentence_embeddings)

Ethics & Safety

Ethics & Limitations

• Domain Specificity: Highly optimized for the Sovereign Knowledge Studio career and protocol definitions; may not generalize well to unrelated domains.
• Language Bias: Trained exclusively on English data.
• Context Length: Performance degrades or truncation occurs beyond 512 tokens.
• Safety: The model has not been aligned with RLHF for conversational safety and should not be used as a generative chat model. Bias might exist in the representations of career fields inherited from the base model.

Citation

@misc{medinamemorysystems2026mesie,
  title={MESIE-768: A Multi-Environment Sovereign Intelligence Engine for Career Protocol Routing},
  author={MedinaMemorySystems},
  year={2026},
  publisher={Hugging Face}
}