NanoTitan-NLI

⚡ NanoTitan-NLI (33M)

NanoTitan-NLI is an ultra-compact, production-grade Natural Language Inference (NLI) and Zero-Shot Text Classification engine. Fine-tuned on the full MultiNLI (392k pairs) corpus in strict 32-bit floating-point precision (FP32), NanoTitan-NLI is designed for high-throughput intent routing, low-latency microservices, and edge deployment where larger transformers create unacceptable latency bottlenecks.

At only 33.3 million parameters (~133 MB disk footprint), NanoTitan-NLI establishes a new performance baseline for sub-50M models, scoring 85.35% on MNLI-Matched and directly outperforming full-size 110M BERT-Base and 66M DistilBERT models while operating with sub-10ms CPU inference latencies.

---

🌟 Key Highlights

• Tier Leader (<50M Parameters): Highest recorded NLI validation score (85.35%) in the sub-50M parameter category.
• Beats Standard BERT-Base: Outperforms bert-base-uncased (84.50%) by +0.85% while consuming 3.3× less memory and disk footprint.
• Zero Quantization Degradation: Trained end-to-end in full FP32 across all 12 transformer encoder layers to preserve exact attention weights.
• Dual-Paradigm Architecture: Operates natively as a 3-class sentence-pair NLI engine (entailment, neutral, contradiction) or as a dynamic Zero-Shot Classifier via premise-hypothesis reformulation.
• Permissive MIT License: Ready for unrestricted commercial and private enterprise deployment.

---

📊 Comprehensive Industry Benchmark Comparison

Evaluated on official test sets across 9 diverse tasks including core NLI entailment, multi-class topic categorization, sentiment polarity, emotion detection, and linguistic pragmatics.

1. Global Performance Matrix vs. Competing Architectures

---

2. Task-by-Task Accuracy Breakdown

• Average Core NLI Accuracy (In-Domain + Cross-Domain): 85.13%
• Average Zero-Shot Generalization (7 Unseen Benchmarks): 63.99%
• Comprehensive 9-Task Overall Average: 68.69%

---

🎯 Production Use Cases

• High-Throughput AI Agent Routing: Categorize user prompts into tool/function execution buckets (["database_retrieval", "code_interpreter", "customer_support", "casual_chat"]) in sub-8ms before routing to expensive frontier LLMs.
• Serverless & Edge Deployments: Deploy on AWS Lambda, Cloudflare Workers, Raspberry Pi, or CPU VPS nodes without GPU dependencies or VRAM exhaustion.
• Real-Time Stream Moderation: High-volume classification of user-generated content, reviews, and support tickets at hundreds of records per second.
• Dynamic Zero-Shot Labeling: Classify unlabelled datasets on the fly simply by defining candidate label arrays at inference time.

---

💻 Quickstart Guide

1. Natural Language Inference (Sentence-Pair Classification)

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "blueprint-ai/NanoTitan-NLI"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

premise = "A soccer match is being played in a packed stadium."
hypothesis = "Athletes are actively competing in a sporting event."

inputs = tokenizer(
    premise, 
    hypothesis, 
    return_tensors="pt", 
    truncation=True, 
    max_length=128
)

with torch.no_grad():
    logits = model(**inputs).logits
    probabilities = torch.softmax(logits, dim=-1)[0]

labels = ["entailment", "neutral", "contradiction"]
for label, score in zip(labels, probabilities):
    print(f"{label:<15}: {score.item() * 100:.2f}%")

2. Zero-Shot Text Classification Pipeline

from transformers import pipeline

classifier = pipeline(
    "zero-shot-classification",
    model="blueprint-ai/NanoTitan-NLI"
)

text = "The Federal Reserve decided to lower interest rates to stimulate market liquidity."
candidate_labels = ["economy & finance", "sports", "entertainment", "healthcare"]

results = classifier(
    text,
    candidate_labels=candidate_labels,
    hypothesis_template="This text is about {}."
)

for label, score in zip(results["labels"], results["scores"]):
    print(f"{label:<20}: {score * 100:.2f}%")

⚙️ Architecture & Training Specifications

Base Checkpoint: microsoft/MiniLM-L12-H384-uncased

Layers: 12 Transformer Encoder Layers

Hidden Dimension: 384

Attention Heads: 12

Total Parameters: 33,360,003

Training Dataset: Multi-Genre Natural Language Inference (nyu-mll/multi_nli) — 392,702 training pairs

Precision: Full FP32 (Strict 32-bit floating point precision)

Optimizer: AdamW (lr=3e-5, weight_decay=0.01, eps=1e-8)

Learning Rate Schedule: Linear Warmup over 3,680 steps (10% warmup)

Total Optimization Steps: 36,816 steps (3 complete epochs)

Batch Size: 32 effective (Batch size 16 × Gradient accumulation 2)

Max Token Length: 128 tokens training limit (Supports up to 512 positional sequence limit)

Hardware: 1× NVIDIA GeForce GTX 1660 Super (6GB VRAM)

Training Wall Time: 1 hour 57 minutes 57 seconds

⚠️ Limitations

Positional Limit: Maximum architectural token capacity is bounded at 512 tokens. Documents exceeding 512 tokens should be truncated or evaluated using sliding-window chunking.

Language Support: Optimized exclusively for English text.

Pragmatic Reasoning: Non-literal language (deep irony, complex sarcasm, and subjective bias) may exhibit lower confidence than direct factual entailment.

📜 License & Citation

This model is licensed under the MIT License.