Fine Tuning Strategies for Language Models and Large Language Models


Language Models and Large Languade Models (LLM) have garnered significant attention in both the public and enterprise sectors due to their proficiency in natural language understanding and generation. These models have shown capacibilities in applications such as direct chat services for general public and natural answers generation as alternative to plain search. Despite their baseline capabilties, the specialized nature of business requires a more precise control and particularly higher accuracy to meet both real business cases as well as specific organizational requirements.

This presentation explores the fine-tuning of Language Models (LMs) and LLMs, starting with the motivations and rationale behind fine-tuning Language Models. The specific demands of highly accurate specialitzed tasks and domains drive the need for fine-tuning to enhance the relevance in specialized contexts.

We introduce general machine learning concepts, such as transfer learning, allowing fine tuning capabilities of Language Models without extensive retraining those models from scratch.

Then, we continue with an overview of actual fine-tuning methods for standard LMs as well as LLMs, and their different use cases during the model development cycle.

The architecture of task heads for fine tuning are examined next. This section details the structure, implementation, and impact of task heads on the adaptability and efficiency of LMs in practical applications.

Then, we introduce the concept and use of adapters for Large Language models, as well as their benefits for resource-efficient alternative. The differences between adapters and traditional task heads are discussed, emphasizing the advantages of adapters in terms of fine-tuning efficiency and model performance.

Furthermore, the impact of adapters in specific applications is examined, showcasing their role in customizing LLMs through learning methods such as DPO and RLHF, with a focus on their integration with Retrieval Augmented Mechanisms.

In conclusion, the presentation provides an overview of fine-tuning methods for LMs and LLMs, highlighting their role in customizing these models for specific business applications. Those insights serve as a basis for further enhance the methodologies in applied LM and LLM.

Background Knowledge:

Audience must be familiar with standard ML and DS concepts.


Kevin Noel is currently Lead AI/ML at Uzabase Japan/US, developing LM / LLM based solutions for Speeda Edge business intelligence. He has more than 12 years experience in Japan in various industries. Previously, he has worked on Ads/Recommendation field with real time personalized Ads solution on Yahoo Japan application.

He also held a principal ML role at the largest Big Data, E-commerce in Japan and has worked with large scale multi-modal data (Tabular, Time series, Japanese NLP, image) through numerous machine learning projects. He has also provided various training on Deep Learning and external talk on applied ML (New York, 2019, ... )... Prior to this, Kevin, with a background in applied stochastic modeling and data mining from Ecole Centrale, held various quantitative roles a BNP Paribas, Bank of America, and ING in Asia/Japan.

Open Data Science




Open Data Science
One Broadway
Cambridge, MA 02142

Privacy Settings
We use cookies to enhance your experience while using our website. If you are using our Services via a browser you can restrict, block or remove cookies through your web browser settings. We also use content and scripts from third parties that may use tracking technologies. You can selectively provide your consent below to allow such third party embeds. For complete information about the cookies we use, data we collect and how we process them, please check our Privacy Policy
Consent to display content from - Youtube
Consent to display content from - Vimeo
Google Maps
Consent to display content from - Google