Research Scientist

Data Science

Core Science

Ads Quality

Founder at Surge AI

Edwin Chen

Surge AI

Head of Gen AI

Manager, Data Operations

Tour/Production Manager

Tour Manager and Touring Drummer

Calvin Harris — Production Manager

Sir Sly — Tour Manager

BROCKHAMPTON — Production Manager

Marketing and Social Media Intern

Music Marketing and Editorial Intern (Zune Marketplace)

Head of Gen AI @ Surge AI

Scott Heiner

Surge is a platform that enables businesses to build and manage their own AI-powered sales teams, automating the sales process and improving customer engagement.

Surge

Surge AI is the world's most powerful data labeling and RLHF platform

Making LLMs reliable is a tough task.

But this is where a lot of the LLM research and development work is focused.

Let's take a look at how LLMs are made reliable:

At Surge AI, we work with the top AI companies to improve LLM reliability.

This effort is essential to enable wider applicability in even higher-stakes domains.

Reliability not only focuses on getting models to output what users want in terms of specifics and quality but also ensuring that no unwanted output (e.g., toxic content) is produced by the model.

A lot of the current efforts to increase reliability focus on ad-hoc approaches and prompt engineering. More recently, there have been more efforts to develop a more systematic framework to improve reliability while training models.

This has led to a lot of interest in red teaming. Red teaming deals with identifying risks in LLMs through adversarial prompting. It has been applied not only to general-purpose LLMs like Claude and ChatGPT but also to more recent code LLMs like Llama Code.

The challenge with red teaming is that, if not done right, it can lead to LLMs over-refusing and potentially leading to a bad user experience. In addition, the reality is that red teaming requires deep expertise in working with LLMs.

We deeply believe that in order to make LLMs safer, useful, and more reliable, comprehensive red teaming is critical. But you don't need to hear this from us. Many large LLM companies have also publicly expressed huge interest in red teaming.

If you are looking for deep expertise in training LLMs and red teaming, reach out to learn how our world-class team can help: surgehq.ai/rlhf

The most powerful LLMs in the world are trained on Surge AI’s RLHF.

Some of these models include code LLMs which are going to be huge for progress in AI!

Let’s look at some of the recent developments in code LLMs:

Foundational Code LLMs - Research indicates that foundational code LLMs could be vital for enabling powerful code understanding and code generation capabilities necessary in more advanced applications of LLMs that range from personalized code assistants to AI-powered debugging tools.

What makes a good code LLM? It’s really about the data quality. As with other types of general-purpose LLMs, high-quality data is key to training code LLMs. Let's look at some recent developments in code LLMs to find out more.

Code Llama - There is a lot we can learn from the recently released code LLM called Code Llama. Code Llama leverages code-specific datasets but there is also a variant called Code Llama - Instruct that leverages proprietary instruction tuning and self-instruct datasets used in Llama 2 to inherit instruction following, helpfulness, and safety properties.

Code Generation Results - Other code LLMs like AlphaCode and StarCode are trained using code only while Code Llama leverages the foundational Llama 2 model. These models are evaluated on description-to-code generation benchmarks. The code-heavy datasets are key to enhanced results in code generation. In fact, even the Code Llama 7B compares or outperforms Llama 70B on Python coding benchmarks and multilingual evaluation.

Red Teaming - Similar to general-purpose LLMs, code LLMs can also benefit from red teaming which involves identifying risks through adversarial prompting in the context of coding. For instance, it can help to prevent the LLM from generating malicious code. While this is beneficial for high-stakes applications, it can result in LLMs over-refusing which might lead to bad user experience in some domains.

It’s clear to see how careful attention to building high-quality datasets can help in training state-of-the-art code LLMs, enabling their utility, and getting desired capabilities.

If you need help with training your code LLMs, red-teaming, or collecting high-quality datasets, reach out to our team: https://lnkd.in/eGiZPbub

Human Evaluation in LLMs

One of the more challenging aspects of LLM research is evaluation.

Here are some of our observations based on past and ongoing research:

At Surge AI, we provide human evaluation for finetuning and RLHF and for both closed and open-source LLMs.

Human evaluation is a challenging and time-consuming endeavor when working with LLMs especially when considering to release or deploy them.

Many automatic and manual approaches have been proposed to properly evaluate LLMs.

For simpler tasks, it’s possible to use out-of-the-box benchmarks or metrics for testing. In some cases, researchers have also relied on LLM-powered evaluators.

However, when evaluating LLMs on more complex tasks and generated outputs, it becomes challenging to use out-of-the-box benchmarks or rely on LLM-based evaluators.

LLM evaluation requires a lot of creativity and robust methodology, requiring deep domain expertise in many cases.

How do LLM researchers perform human evaluation and what are some considerations?

Given the vast range of use cases LLMs cover, researchers and practitioners evaluate different aspects of LLMs like safety, helpfulness, creativity, and many other characteristics.

This is extremely difficult to achieve with standardized evaluations. For this reason, human evaluation has become more widespread with LLMs.

Human evaluation entails assessing the performance of LLMs through human labelers which comes with its own challenges like consistent labeling and difficulty of labeling for complex tasks.

Specifically, human evaluation can be unreliable due to factors such as:

• The prompts being evaluated may not be comprehensive enough to cover all possible scenarios

• Raters may also have different preferences, which can lead to different ratings for the same generation

• The evaluation guidelines may be open to interpretation, which can lead to evaluators giving different ratings for the same generation

• It can be difficult to compare generations of text generation models due to the interface or simply the type of generations these LLMs are producing

We work with top LLM companies to help with all of these human evaluation challenges and more.

If you need help with carrying out high-quality human evaluation on your LLMs, reach out to our team here: https://lnkd.in/eGiZPbub

#machinelearning #deeplearning #ai

What is SFT data and what role does it play in state-of-the-art LLMs?

Supervised finetuning (SFT) in the context of RLHF deals with further tuning an initial language model using demonstration data.

At Surge AI, we provide SFT data for top LLM teams to finetune their LLMs. 

Here is what we have observed:

SFT data typically involves collecting demonstration data including prompts and in-depth responses written by human annotators demonstrating how the model should respond to the prompt. Specifically, You take a set of commands and obtain human-written responses for each.

The SFT training dataset consists of <prompt, ideal generation> pairs used to finetune the pre-trained LLM to output human-like responses. So let’s say you are aligning an LLM-powered dialogue system then you need to collect dialogue-style instructions/responses data that cater to that use case. Similarly, as shown in the figure, if you want high-quality code generation capabilities you can also provide instruction + written responses as part of the SFT data.

This leads to the first important component, also referred to as supervised policy, for training an RLHF LLM. But why go through all this process when training RLHF LLMs? The core idea of SFT is to provide a high-quality initialization for the RLHF process. It’s widely applied by some of the most advanced closed and open-sourced LLMs.

To make this work, you need to collect lots of demonstration data but the challenge is collecting high-quality and diverse demonstration data at scale. SFT data can be written by different annotators and can incorporate a lot of noise as response quality and style can vary from annotator to annotator. Controlling for this is key.

According to reported insights, you will need to collect thousands of examples to ensure you are tuning a high-quality LLM. SFT data helps to improve target areas that allow steering the LLM better to your needs.

We can help with your SFT data needs! If you need help with collecting high-quality preference or SFT data, reach out to our team here: https://surgehq.ai/rlhf

RLHF enables some of the most powerful LLMs today.

One of the key components in RLHF LLMs is human preference data which makes it possible to align LLMs with human preferences.

Here are some key insights:

The reward model fundamentally enables tuning an RLHF LLM to align with human preferences. So naturally, high-quality comparison data is key to making it work.

We at Surge AI work with the top LLM teams to provide annotations for both SFT and reward models.

Here are the typical steps involved in training the human feedback reward model:
• you need to form a new set of commands
• use an LLM to generate responses
• then ask Surgers (our human LLM evaluators) to score/rank these responses
• the next step is to use train the reward model on these human feedback scores

Keep in mind that the goal of the reward model is not to model knowledge, it is used to enable the LLM to effectively model human preferences and interactivity.

As shown in many research papers, training a strong reward model in RLHF is challenging but, if done right, can lead to more aligned models and improve helpfulness and safety in LLMs.

These components power some of the most advanced LLMs like ChatGPT and Claude to perform useful tasks for real-world applications like code generation.

Key insights from research on RLHF LLMs and reward models indicate that: 1) the reward model relies on high-quality human data annotations, and 2) to ensure the output quality of RLHF LLMs, it’s key to collect this high-quality human preference data at scale and potentially find strategies to keep collecting even higher quality data.

If you are curious about how human preference data can enable your LLMs or need help collecting high-quality preference or SFT data, reach out to our team here: https://surgehq.ai/rlhf

Photo	Title	Links
	Edwin Chen CEO
	Scott Heiner Head of Gen AI

AI:SURGEHQ

Surge

Preview

Edwin Chen

CEO

Scott Heiner

Head of Gen AI