That document has an incorrect definition of prompt injection: it says "Prompt injection attacks are attempts to circumvent content restrictions to produce particular outputs" – but that's not prompt injection, that's jailbreaking
@simonw
-
SQuAD Explorer: Benchmark for Machine Reading Comprehension
By
–
Oh yeah, this looks good https://
rajpurkar.github.io/SQuAD-explorer/ -
Reka AI Advanced Chat Capabilities Discovered
By
–
I spotted that ability for Reka too https://
docs.reka.ai/guides/001-cha
t.html#advanced-chat-usage
… -
ChatGPT Language Bias: Nigerian English RLHF Hypothesis
By
–
I wonder how we could prove this one way or another? The idea that ChatGPT's language usage is biased towards formal Nigerian English based on RLHF outsourcing is pretty irresistible, I want it to be true!
-
Basic Reranker Model Example Input Output Guide
By
–
Suggestion: include a really basic example of what a reranker model can do, with example input and output
-
Meta’s Instruct and Non-Instruct Models Available on Replicate
By
–
How do you mean? Replicate have both the instruct and the non-instruct models linked here: https://
replicate.com/meta -
Llama 3 API Deployment Status Check Among Users
By
–
Anyone else got an API running serving Llama 3 yet?
-
Llama 3 Models Now Available on Replicate Platform
By
–
You can run prompts to rough the new Llama 3 models on Replicate: https://
replicate.com/meta/meta-llam
a-3-70b-instruct
… for 70b-instruct and https://
replicate.com/meta/meta-llam
a-3-8b-instruct
… for 8b-instruct -
ChatGPT vocabulary bias traced to Nigerian RLHF annotation workers
By
–
Love this theory by @alexhern that the reason ChatGPT uses words like "delve" a lot is that OpenAI outsource a lot of their RLHF annotation to workers in Nigeria, and Nigerian formal English has a slightly different vocabulary