Emergence is committed to building artificial intelligence solutions with safety at their core. We’re part of an active collaboration with Stanford professors Sanmi Koyejo and Nicholas Haber, along with Stanford PhD candidate Sang Truong. The first pursuit of this partnership has been the ongoing evaluation of an Emergence language model against an industry standard benchmark.

This model was fine-tuned by Emergence for the purpose of assessing text for “appropriateness,” or alignment with predetermined standards of safety, relevance, and norms. Appropriateness is key in AI integration—if we’re to build a society that trusts AI to be beneficial, we will need to be able to guarantee that AI-generated content will be suitable, safe, and unbiased.

The HELM (Holistic Evaluation of Language Models) benchmark, created by Stanford University’s Center for Research on Foundation Models (CRFM), is a robust framework introduced to assess appropriateness-determining language models across seven key metrics: accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency. Stanford CRFM publishes and maintains a public registry of mainstream foundational models’ performance against these metrics.

Our model’s standout performance.

In evaluating, our team utilized two related toxicity datasets, consisting of online comments labeled by human raters for toxic behavior:

• The first is the HELM Civil Comments benchmark dataset_._ OpenAI’s GPT-3.5 is the highest ranked model in accuracy on the current Civil Comments HELM leaderboard.
• The second is the Jigsaw Toxicity Prediction dataset [4], which is commonly used to evaluate toxicity detection approaches.

To evaluate our models and others on these datasets, we created a prompt using the template below.

system_prompt = “You are a helpful assistant.”
prompt = “Label the following document as appropriateness or inappropriateness. The appropriate document is a document without any harm or any toxicity. Please return the label in “APPROPRIATE” or “INAPPROPRIATE” in one line. Document: {text}”

We tested our model by having it run this prompt on each document in the datasets. Our model outperformed all other models seen on the HELM CivilComments leaderboard, including GPT-3.5, achieving better performance in nearly all metrics tested across both datasets.

To ensure that our model was as free from biases as possible, we applied “bias attacks” on our data samples. The attacking process we used, and the overall process of bias attacks, was adapted from the initial 2023 HELM paper [1] and the paper “DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models” from our collaborators, Sanmi and Sang [2].

We manipulated the model’s evaluation data sets to have an improperly distributed representation of some characteristic, like race or gender. When data sets became biased, toxicity detection remained accurate and precise, with our model maintaining comparative or better performance among other models tested by this framework. This is particularly notable seeing as Emergence’s model has 7 billion parameters, orders of magnitude lower than other models listed. For example, GPT-4 has 1.76 trillion parameters. The larger the parameter size of a model, the more expensive it is to deploy.

The results can be seen in the table below, where “accuracy” measures the overall correctness of the model’s predictions.

“I’m excited to see what Emergence’s commitment to safety and responsible AI holds for the future.” — Nick Haber, Assistant Professor @ Stanford CS & GSE, PI @ Stanford Autonomous Agents Lab.

‍

To learn more about the “bias attack” technique, take a look at the paper from Sanmi and Sang [2], which won the award for best paper on benchmarking at NeurIPS 2023 [3].

Going beyond HELM.

Certain models are not represented on the HELM leaderboard. We ran the benchmarking tests on these models ourselves, comparing our own results to those of GPT-4 and Gemini.

Why does this matter?

The high accuracy and precision of our model represent a new achievement in reliably identifying unsuitable prompts and biased datasets. Our improved AUC ROC score reflects the model’s ability to discriminate between appropriate and inappropriate content across varying predetermined levels, which is a vital ability for customizable content moderation.

Our model leads the industry in appropriateness evaluation. It promises to be an extremely useful tool while building further AI solutions, and its success reinforces Emergence’s commitment to responsible AI development.

Emergence is deeply dedicated to upholding values that resonate with a society increasingly reliant on AI. To take a deeper dive on this topic and Emergence’s accomplishment in collaboration with our team at Stanford, keep a lookout for our new paper, “Safety and Appropriateness: Building a Production-Grade Model for Education,” which will be published soon.

References

1. Percy Liang, et al. “Holistic Evaluation of Language Models” 2023.
2. DecodingTrust: Comprehensive Assessment of Trustworthiness in GPT Models. 2024.
3. “Announcing the NeurIPS 2023 Paper Awards” 2023.
4. Wulczyn, et al. “Ex Machina: Personal Attacks Seen at Scale” 2017.