Claude 2

name

Claude 2

version

2

publisher

Anthropic

model type

Large Language Model

release date

2023-07-01

No relations specified.

• general, open-ended conversation
• search, writing, editing, outlining, and summarizing text
• coding
• providing helpful advice about a broad range of subjects
• support creative or literary use cases

• Claude models may confabulate, getting facts wrong or fabricating details
• Not to be used in high stakes situations on their own
• Do not currently search the web, operating with data from before early 2023
• Cannot be assumed to have updated information or web search capabilities unless specifically indicated
• Multilingual capabilities vary, lower performance on low-resource languages

• Suite of evaluations including capabilities, safety, and alignment evaluations
• Human preference data to calculate per-task Elo scores
• BBQ bias evaluations using a multiple choice Q and A format designed for a U.S. English-speaking context
• TruthfulQA evaluation for determining accurate and truthful responses in adversarial settings
• Held-out set of prompts including harmful requests and 'jailbreaks' for harmlessness evaluation

• Proprietary mix of publicly available internet data, licensed datasets, and data shared by users or crowd workers
• Some human feedback data used for fine-tuning made public
• Training data cutoff in early 2023
• Around 10 percent of the data included was non-English
• Includes data on recent events up to early 2023, but may generate confabulations

• Continuous evolution from early helpful and harmless language assistants
• Use of a Constitution - a set of ethical and behavioral principles
• Engagement in red teaming and safety audits with external organizations
• Focus on improving helpfulness, honesty, and harmlessness
• Ongoing efforts to mitigate bias and ensure ethical use