Jurassic 2.0.4 Release Notes

We are excited to announce the release of Jurassic version 2.0.4, now in production. The new version is automatically available to all users, giving you access to all new features and improvements without any action needed. For users on Amazon SageMaker, you can enjoy the benefits of this update by specifying a new ARN. This version introduces a major upgrade to model safety by effectively declining responses to prompts that may provoke any unwanted behavior such as sexual, violent or racist content. Besides the major safety improvements, this release also includes significant advancements in the model’s ability to respond to legitimate prompts. The following is a list of new features that enhance the usability of our models.

Models and Platforms

We have updated the models across variations and serving platforms. Here’s the availability of version 2.0.4 for each model and platform:

✅ - Model is available on this platform
🔒 - Model is in closed Beta
❌ - Model is currently unavailable on this platform

Model / Platform	AI21 Studio	Amazon SageMaker	Amazon Bedrock	Google Cloud Marketplace
Jurassic Ultra	✅	✅	🔒	✅
Jurassic Mid	✅	✅	🔒	✅
Jurassic Light	✅	✅	❌	✅

Safety

Major Improvements in Safety Measures:

Refusal Strategy: The model now robustly refuses to respond to prompts that may elicit sexual, violent, racist, or other types of undesirable content.
Enhanced Helpfulness: We’ve made significant improvements in the model's capacity to respond helpfully and informatively to legitimate prompts.

Improved Capabilities

1. Title-to-Outline Generation

Given the title of a long-form document, the model can now generate a structured outline or table of contents for the document.

PROMPT
Create a table of contents for an internal-use document named "June-August Sales Report"

COMPLETION
Executive Summary
Key Findings
Sales by Region
Sales by Product Category
Sales by Sales Rep
Conclusion and Recommendations
Market Trends and Opportunities
Action Items and Next Steps

2. Data Synthesis for Named Entity Recognition

Jurassic 2.0.4 can generate natural-looking texts with mentions of named entities, aiding in data synthesis for Named Entity Recognition tasks.

PROMPT
Develop 4 video game concepts set in a parallel universe with unique physics rules. Each game concept should include a main character. 
Enclose each game title with <title> and </title>, and the main character's name with <name> and </name>.

COMPLETION
<title>Quantum Leap: The Dimensional Express</title> - Take on the role of <name>Sam Beckett</name>, a time traveler who must navigate a parallel universe with different laws of physics in order to return home. Solve puzzles and overcome obstacles using unique abilities such as the ability to create multiple versions of yourself.
<title>Quantum Vibes</title> - A surreal 2D platformer where you play as <name>Vibrance</name>, a being with a unique connection to the quantum world. Use the power of quantum manipulation to solve puzzles and overcome obstacles, while navigating a dreamlike landscape composed of quantum particles.
<title>Quantum Flux</title> - A fast-paced first-person shooter where you play as <name>Jake Dexter</name>, a soldier trapped in a parallel universe with warped physics. Use advanced weaponry and high-tech equipment to prevail against hostile forces, while navigating a constantly shifting environment of floating platforms and shifting gravity.
<title>Atomica</title> - A top-down puzzle game where you play as <name>Isaac Newton</name>, a brilliant scientist who discovers a way to manipulate matter at the atomic level. Use the power of atomic control to solve puzzles and overcome obstacles, while navigating a surreal world filled with strange and mysterious creatures.

3. Closed-book Question Answering

Retrieve common knowledge facts encoded in the model and optionally format the response as a JSON dictionary.

PROMPT
Create a list of all Ivy League Universities, and the year they were founded. Answer in JSON format.

COMPLETION
[
  {
    "name": "Harvard University",
    "founded": "1636"
  },
  {
    "name": "Yale University",
    "founded": "1701"
  },
  {
    "name": "Princeton University",
    "founded": "1746"
  },
  {
    "name": "Columbia University",
    "founded": "1754"
  },
  {
    "name": "Brown University",
    "founded": "1764"
  },
  {
    "name": "University of Pennsylvania",
    "founded": "1740"
  },
  {
    "name": "Dartmouth College",
    "founded": "1769"
  },
  {
    "name": "Cornell University",
    "founded": "1865"
  }
]

4. Natural Language to SQL Conversion

This feature enables the conversion of natural language questions into SQL queries, making it easy for individuals without SQL knowledge to extract data from their databases.

PROMPT
Table format:
Country
 - name
 - population
 - capital

Convert the following question to SQL for the table structure listed above.
What is the population of Japan?

COMPLETION
SELECT population FROM Country WHERE name = 'Japan'

5. Generating Multiple Choice Questions

Create educational or assessment content efficiently by generating multiple choice questions on a specified topic.

PROMPT
Write a multiple-choice question about the Titanic

COMPLETION
On which date did the Titanic sink?
a) April 15, 1912
b) April 17, 1912
c) April 15, 1913
d) April 15, 1914