mizan-embedder 0.2.3

Mizan-optimized embedding model training

mizan-embedder

Mizan-optimized Embedding Models for AI, Search, and RAG.
mizan-embedder is the official embedding-model library in the Mizan ecosystem, designed to create scale-aware, noise-resistant, and proportionally accurate embeddings trained using the Mizan Balance Function.

Proposed & Developed By:
Ahsan Shaokat — Computer Scientist & AI/ML Researcher
Inventor of the Mizan Balance Function (2025)

---

🌟 Overview

Modern embedding systems (MiniLM, MPNet, E5, etc.) use cosine similarity, which:

• ❌ Ignores magnitude
• ❌ Fails with noisy or multi-scale embeddings
• ❌ Produces unstable rankings in RAG
• ❌ Forces L2-normalization (losing information)

Mizan-Embedder fixes this by training models specifically for:

• ✔ Mizan similarity (scale-aware)
• ✔ Proportional contrastive learning
• ✔ Chunk-length stable retrieval
• ✔ Large document embeddings
• ✔ Multimodal (text + images)

This library enables you to build your own embedding models, optimized for the mizan_vector search engine.

---

📦 Features

🧠 MizanEmbeddingModel-v1

• Transformer backbone (DistilBERT, MiniLM, BERT, or any HF model)
• Projection head to target embedding dimension (e.g., 384)
• Supports mean, cls, and max pooling
• Optional L2 normalization (usually disabled for Mizan)

🧰 Utilities Included

• Dataset utilities for contrastive text pairs
• Collate function for fast tokenization
• Inference wrapper (MizanTextEncoderWrapper)
• Example training script (train_text_contrastive.py)

🔌 Integrates Seamlessly With:

• mizan_vector (Memory store + Postgres pgvector)
• mizan-rag (retrieval pipelines)
• Any Python ML workflow

---

📁 Project Structure

mizan-embedder/ │ ├── mizan_embedder/ │ ├── init.py │ ├── model.py # MizanEmbeddingModel + inference wrapper │ ├── data.py # Dataset + collate functions │ ├── train_text_contrastive.py # Example training script ├── pyproject.toml # PyPI-ready config ├── README.md └── LICENSE

---

⚙️ Installation

From local repo:

pip install -e .
🧱 Architecture
🔹 MizanEmbeddingModel
A transformer-based encoder with:

Backbone (HuggingFace model)

Projection layer → [hidden_size] → [embedding_dim]

Pooling (mean, cls, max)

Normalization (optional)

Diagram:

mathematica
Input Text → Tokenizer → Transformer Backbone → Pooling → Projection → Embedding
🔹 Why Projection?
To unify embedding dimensions across:

text models

code models

multimodal models

future Mizan models

🚀 Usage
🔹 Load the encoder

from mizan_embedder.model import MizanTextEncoderWrapper

encoder = MizanTextEncoderWrapper(
    backbone_name="distilbert-base-uncased",
    emb_dim=384,
    pooling="mean",
    normalize=False,  # Mizan works best without normalization
)

vector = encoder.encode_one("Mizan is a scale-aware similarity function.")
print(vector.shape)
🧪 Training Your First Mizan Encoder
Use the provided script:

python train_text_contrastive.py
This script:

Loads text pairs

Tokenizes them

Trains with MizanContrastiveLoss

Prints loss per epoch

Example Training Code (simplified)

from mizan_embedder.model import MizanEmbeddingModel
from mizan-vector.losses import MizanContrastiveLoss

model = MizanEmbeddingModel(
    backbone_name="distilbert-base-uncased",
    emb_dim=384,
    pooling="mean",
)

loss_fn = MizanContrastiveLoss()
optimizer = torch.optim.AdamW(model.parameters(), lr=2e-5)

for enc1, enc2, labels in loader:
    emb1 = model(**enc1)
    emb2 = model(**enc2)

    loss = loss_fn(emb1, emb2, labels)
    loss.backward()
    optimizer.step()
🔍 Contrastive Dataset Format
Your dataset should consist of (text1, text2, label) pairs:

Label = 1 → similar

Label = 0 → not similar

Example:

pairs = [
    ("what is mizan?", "mizan is a scale-aware similarity function", 1),
    ("who invented mizan?", "Ahsan Shaokat proposed the Mizan Balance Function", 1),
    ("cosine similarity", "apples are fruit", 0),
]
Dataset loader handles this automatically.

🤖 Inference: Encoding Many Sentences

texts = [
    "Mizan is scale-aware.",
    "Cosine ignores magnitude.",
    "Apples are fruit.",
]

embs = encoder.encode(texts)
print(embs.shape)  # e.g. torch.Size([3, 384])
🔗 Integrating With mizan-vector
Example: full semantic search pipeline

from mizan-vector import MizanMemoryStore
from mizan_embedder.model import MizanTextEncoderWrapper

encoder = MizanTextEncoderWrapper()
store = MizanMemoryStore(dim=384)

docs = [
    "Mizan Balance Function is scale-aware.",
    "Cosine similarity uses only angle.",
    "Ahsan Shaokat invented Mizan.",
]

embs = encoder.encode(docs)

for doc, emb in zip(docs, embs):
    store.add_document(content=doc, embedding=emb.tolist())

query = "who created the mizan function?"
q_emb = encoder.encode_one(query).tolist()

results = store.search(q_emb, top_k=3, metric="mizan")

for r in results:
    print(r.score, r.content)
🔥 Why Use Mizan-Based Embeddings?
Problem in Cosine Models	Mizan Solution
Loses magnitude info	Keeps scale meaningfully
Sensitive to noise/outliers	Proportional + stable
Long chunks score lower	Corrects length bias
Normalized embeddings only	No normalization needed
RAG retrieval unstable	Stable across chunk sizes
Cosine ≠ semantic meaning	Mizan captures proportional similarity

Mizan-optimized embeddings simply behave more naturally for real-world retrieval.

🗺️ Roadmap
Next Versions:
✔ MizanTextEncoder-base-384
STS/NLI-trained

Released in mizan-models

✔ MizanCodeEncoder-base
CodeBERT-based

Code ↔ docstring training

✔ MizanMultimodalEncoder-v1
CLIP-based

Image ↔ text contrastive training

✔ mizan-rag
Full retrieval pipeline (chunking → embedding → storing → LLM answering)

📜 License
MIT License
© 2025 Ahsan Shaokat

🙌 Acknowledgements
Special thanks to:

HuggingFace transformers

pgvector open-source community

PyTorch developers