d
WE ARE LINGO
2nd floor, Garuda BHIVE, BMTC Complex,
Old Madiwala, Kuvempu Nagar,
Stage 2, BTM Layout, Bengaluru,
Karnataka – 560068. India.
SandLogic Lexicons

A Hub for Optimized Large Language Models

SandLogic Lexicons, the go-to repository for cutting-edge Large Language Models (LLMs) optimized for deployment on various hardware platforms. Our collection bridges the gap between the latest AI research and practical applications, making advanced natural language processing capabilities accessible to all.

Github Repository
Hugging Face

SandLogic Lexicons offers a diverse array of pre-trained and fine-tuned LLMs, carefully optimized to deliver top-tier performance and efficiency. Our models are designed to operate smoothly on both GPU and CPU architectures, broadening access to advanced AI capabilities. Our mission is to make LLM technology widely accessible to developers, researchers, and AI enthusiasts. 

Key Features

Optimized Performance

Every model in our Lexicons undergoes a stringent optimization process to ensure maximum performance across a range of hardware configurations. Whether you’re using high-performance GPUs or standard CPUs, our models are crafted to deliver outstanding results without sacrificing efficiency. 

Advanced Quantization Techniques

We utilize state-of-the-art quantization methods to minimize model size and boost inference speed, maintaining high accuracy levels. Our quantization techniques include: 

  • Legacy methods such as Q5_K_M and Q4_K_M
  • Innovative approaches like IQ4_XS imatrix quants

These techniques enable the deployment of powerful models on devices with limited resources, making edge AI applications feasible.

Benchmark-Validated Performance

Each model in SandLogic Lexicons is thoroughly evaluated against industry-standard benchmarks, ensuring robust performance across various domains: 

  • HellaSwag: Tests a model's ability to apply commonsense reasoning to complete scenarios plausibly.
  • MMLU (Massive Multitask Language Understanding): A comprehensive benchmark that spans 57 subjects, assessing a model’s multitask capabilities across STEM, humanities, and social sciences.
  • CodeXGLUE Code2Text: Evaluates a model's ability to generate natural language descriptions from source code.
  • Winogrande: A large-scale dataset designed to test commonsense reasoning, derived from the Winograd Schema Challenge.
  • MedMCQA: Focuses on evaluating medical knowledge and reasoning through a multiple-choice question format.
  • MMLU Clinical Knowledge: A specialized subset of MMLU that assesses a model's understanding of clinical knowledge and medical concepts.

These benchmarks provide a thorough evaluation of our models, covering general knowledge, reasoning, coding skills, and specialized medical understanding. 

Lexicon Benchmarks

Small Language Models (SLMs) Finetuning

Model Configuration Training Tokens Infrastructure Original Model SandLogic Model
8 Biliion
3 Trillion
8xA100
21 days
14 days
7 Billion
3 Trillion
8xA100
16 days
12 days
3 Billion
1 Trillion
4xA100
8 days
3 days

With Us

  • Deploy models across variety of hardwares
  • Save on OpEx
  • Ubiquitous deployment of a trained model - Servers, PCs, Mobiles, Wearables (watches)
  • Data security baked into the framework.

Inference Time

Model Configuration Infrastructure (CPU, GPU) Inference Time Taken (seconds) for 400 tokens SandLogic Inference Time Taken (seconds) for 400 tokens
8B
GPU (T4)
11
5
GPU (A100)
7
4
CPU (Intel Xeon 12 cores – 100 GB RAM)
35
28
7B
GPU (T4)
9
4
GPU (A100)
7
3
CPU (Intel Xeon 12 cores – 100 GB RAM)
27
21
3B
GPU (T4)
7
3
GPU (A100)
4
2
CPU (Intel Xeon 12 cores – 100 GB RAM)
15
11

See Our Current Models in Repository

Our repository continues to expand, providing a growing selection of optimized models for diverse applications and hardware platforms. 

Visit Our Github Repository