Generative Large Language Models

AppTek offers state-of-the-art research and development for generative pre-trained transformer (GPT) large language models (LLMs) for generation of fluent human-like text.

Home / Technology / Generative Large Language Models

Specializing in the Research, Development and Deployment of Responsible High-Performance Large Language Models

Generative Large Language Models (LLMs) are advanced systems designed to understand and generate human-like text based on the patterns and structures learned from vast datasets. Models are trained on diverse data sources which enable them to perform a wide range of language-based tasks such as composing essays, answering questions or creating dialogue, amongst many other applications. By leveraging AppTek's data services combined with sophisticated machine learning techniques, including supervised, unsupervised, and reinforcement learning, enterprises can cost-effectively harness and deploy LLMs to generate coherent and contextually relevant text for tasks such as content creation, customer service, education, and more.

AppTek Large Language Modeling experience and expertise:

  • Deep experience in LLM-based training software including Megatron, HF trainer, Nemo, Pytorch FSDP, gpt-neox,Waferscale accelerator, etc.
  • Cluster management and automation of tasks such as script/data/model distribution, distributed training, automatic monitoring, checkpointing etc.
  • Foundational model training with continued pre-training on 7B, 13B, 70B models on 4B in-domain tokens
  • Built-in integration with ASR, MT, video classifiers and object recognition tools
  • Reward modeling and reinforcement learning from human feedback (RLHF)
  • Retrieval-augmented generation including vector database creation, document cleaning and embedding, semantic search, adapted Instructional Fine Tuning (IFT)
  • Cascaded machine translation and domain adaptation with enhancement of performance on low-resource languages
  • Deployment of models such as through the vLLM/TGI inference server, or a simple prototype UI to production-level UI
a woman using the workbench platform

AppTek Data Services for Multilingual LLM Deployment

AppTek's distributed workforce develop and delivers the highest quality data for the development and deployment of bespoke LLM's based on a specific use case or client domain.  AppTek's data services team develops multilingual prompt-pairs for LLMs with training methodology including:

  • Retrieval-Augmented Generation (RAG): Combines retrieval-based methods with generative models to enhance responses by grounding them in specific, retrieved information from a large corpus.
  • Instruction Fine-Tuning (IFT): Fine-tunes a pre-trained LLM on a dataset of task-specific instructions and examples to improve its ability to understand and follow explicit instructions.
  • Self-Supervised Learning: Trains the model by predicting parts of the input data from other parts, using large amounts of unlabeled data to improve generalization and robustness.
  • Reinforcement Learning from Human Feedback (RLHF): Utilizes human feedback to adjust the model's parameters, aligning its outputs with human preferences and ethical standards.
  • Transfer Learning: Adapts a model trained on one task to perform well on another, leveraging pre-existing knowledge to quickly adjust to new tasks with limited data.
  • Few-Shot Learning: Enables the model to perform new tasks with only a few examples by generalizing from minimal task-specific data.

AppTek Specialized Research and Development for LLMs

Domain-Specific LLMs and RAG

AppTek offers experience in pre-training instruction finetuning (IFT) and retrieval augmentation (RAG) that yield strong models which are small in size and comparable in performance compared to much larger models.

Faster Foundational Model Training

Research is underway for count-motivated word vector initialization as well as branch optimization, which both have shown to offer a performance boost for foundation model deployment at the early stage of training.

Smaller Memory During Training

By systematically checking the gains and losses of replacing the adaptive learning rate optimization algorithm, scientists have found improvements in GPU memory resource optimization which could change common LLM training practices.

Reward Modeling and Preference Optimization

By focusing on decision boundaries between similarly ranked completions (going over adjacent completion pairs versus going over all completion pairs) and introducing a margin term for optimization, modern research shows improvements in reward modeling and preference optimization.

Hallucination Detection

To detect hallucinations, scientists are discovering ways to define a quantity for "familiarity of context" and detect when and where hallucinations happen by thresholding this quantity. The process includes collecting context vector statistics during training, maintaining an unsupervised model of their distributions, and querying this unsupervised model during testing.

Automatic Evaluation Metrics

The team has found new ways in R&D for summarization evaluation to calculate the quantity p (original_document | summary) using a strong external LLM that can in turn capture how much information from the original document is retained in the summary automatically for evaluation purposes.

Have questions? Contact AppTek today!

Contact Us
AI and ML Technologies to Bridge the Language Gap
Find us on Social Media:

AppTek is a global leader in artificial intelligence (AI) and machine learning (ML) technologies for automatic speech recognition (ASR), neural machine translation (NMT), natural language processing/understanding (NLP/U) and text-to-speech (TTS) technologies. The AppTek platform delivers industry-leading solutions for organizations across a breadth of global markets such as media and entertainment, call centers, government, enterprise business, and more. Built by scientists and research engineers who are recognized among the best in the world, AppTek’s solutions cover a wide array of languages/ dialects, channels, domains and demographics.

Copyright 2021 AppTek    |    Privacy Policy      |       Terms of Service     |      Cookie Policy