Aletheia MiniLM-L12-v2 (zafaralsathiq/aletheia-minilm-l12)

This is a 12-layer dense sentence embedding model mapping natural language sentences and scientific abstracts into a 384-dimensional dense vector space.

It is optimized for Aletheia Search, a Zero-Hub scientific discovery and synthesis engine that couples dense neural text embeddings with Kim-Vu $d$-regular expander graphs to eliminate citation mega-hub bias and discover unbridged cross-domain research literature.

---

Model Details

• Model Type: Sentence Transformer (12-layer BERT encoder)
• Base Architecture: MiniLM-L12 (384 hidden dimensions, 12 attention heads, 1536 intermediate size)
• Output Dimensionality: 384 dimensions
• Max Sequence Length: 512 tokens
• Pooling Strategy: Mean pooling over attention-masked token embeddings
• Normalization: L2 unit norm ($|\vec{v}|_2 = 1.0$)
• Weights Format: Zero-copy binary model.safetensors (128 MB)

---

Intended Use

1. Scientific Literature Retrieval & Clustering: Semantic indexing of peer-reviewed papers, preprints, and clinical trial abstracts.
2. Topological Expander Routing: Feeding dense affinity scores into degree-regularized graph routers (e.g. Aletheia Kim-Vu coupling).
3. Literature-Based Discovery (LBD): Identifying cross-domain unbridged structural voids and Swanson 2-hop triads ($A \to B \to C$).
4. General Text Similarity & Semantic Search: Drop-in replacement for 6-layer all-MiniLM-L6-v2 with greater semantic depth and representation capacity.

---

Usage

1. Using sentence-transformers (Recommended)

pip install sentence-transformers

from sentence_transformers import SentenceTransformer

# Load model directly from Hugging Face
model = SentenceTransformer("zafaralsathiq/aletheia-minilm-l12")

sentences = [
    "Targeting KRAS G12D mutations with small molecule inhibitors",
    "Novel therapeutic avenues for pancreatic ductal adenocarcinoma",
    "Photosynthetic carbon fixation efficiency in marine red macroalgae",
]

# Generate L2-normalized embeddings
embeddings = model.encode(sentences, normalize_embeddings=True)
print("Embedding shape:", embeddings.shape)  # (3, 384)

# Compute cosine similarity via dot product
similarity = embeddings[0] @ embeddings[1]
print(f"Cosine Similarity (Related): {similarity:.4f}")

---

2. Using Hugging Face transformers & PyTorch

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

tokenizer = AutoTokenizer.from_pretrained("zafaralsathiq/aletheia-minilm-l12")
model = AutoModel.from_pretrained("zafaralsathiq/aletheia-minilm-l12")
model.eval()

texts = ["Targeting KRAS G12D mutations with small molecule inhibitors"]

# Tokenize
encoded = tokenizer(texts, padding=True, truncation=True, max_length=512, return_tensors="pt")

with torch.no_grad():
    outputs = model(**encoded)
    token_embeddings = outputs[0]
    
    # Mean pooling
    mask = encoded["attention_mask"].unsqueeze(-1).expand(token_embeddings.size()).float()
    sum_embeddings = torch.sum(token_embeddings * mask, 1)
    sum_mask = torch.clamp(mask.sum(1), min=1e-9)
    pooled = sum_embeddings / sum_mask
    
    # L2 Unit Normalization
    embeddings = F.normalize(pooled, p=2, dim=1)

print("Vectors:", embeddings.shape)

---

3. Using Native Rust with Candle

The repository provides zero-copy model.safetensors and tokenizer.json for native execution in Rust without Python:

cargo add candle-core candle-transformers tokenizers safetensors

---

Empirical Verification

Measured on domain scientific literature pairs:

---

Citation & Attribution

If you use this model in your research, cite:

@software{aletheia_search_2026,
  author = {Zafar Al Sathiq},
  title = {Aletheia Search: Zero-Hub Scientific Discovery Engine},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/zafaralsathiq/aletheia-minilm-l12}
}