library(ellmer)Can we trust the outputs that we are receiving from GenAI products?
Well trust is built in part, on transparency.
The reason more traditional data science and analytics work found its way into mainstream decision making is because it was transparent and trusted. Most serious data scientists I speak to spend a lot of time worrying about reproducibility. From a scientific and computational perspective, this means their analysis be understood and reproduced by others and achieve the same results.
Why is this important?
Reproducibility enables us to understand, criticize and improve on the work of others. Without it, we are just making one flashy claim after another.
While the promises of GenAI are many - how can it ever be used reliably in decision making if we cannot understand, reproduce, critique and improve on each other’s work?
This is a challenge for Generative AI as its underlying mechanics are non-deterministic. Meaning for the exact same inputs, it will return different unpredictable outputs.
Chat Completions are non-deterministic by default (which means model outputs may differ from request to request).
- https://platform.openai.com/docs/advanced-usage
How can this be addressed?
My challenge for users of GenAI is to meet the same standards as a cooking recipe book: Write down the exact steps you followed so someone else can follow them and get the same results.
Practically speaking:
- If using a large language model (LLM), interact with it using code via an API. This will allow others to rerun it:
- With the exact same prompt
- With the exact same context (not whatever messed up history you have accumulated in your browser)
- With the exact same model
- Set a seed. OpenAI (and possibly others) are introducing ‘seeds’ which allow you “some control towards deterministic outputs”.
- Share the code and the output (if you can) so others can replicate your work.
Do seeds work?
Sort of.
Disappointingly, this feature is only in Beta and apparently only supported for models: gpt-4-1106-preview and gpt-3.5-turbo-1106. It also doesn’t really work consistently, but I’ve found it does help a lot.
Testing reproducible LLM interactions
Below I will use the {ellmer} R package to generate an API call to an OpenAI model with a basic prompt:
“Generate a short excerpt of news about a journey to Mars”
I will do this twice without setting a seed value, and then I will repeat it with a seed value set and roughly compare the output.
First I will create a function to send a basic prompt to an OpenAI LLM.
generate_response <- function(prompt,
system_message,
model,
seed = NULL) {
print("LLM RESPONSE ==========================")
chat <- ellmer::chat_openai(
model = model,
system_prompt = system_message,
turns = NULL,
seed = seed,
echo = "text"
)
chat$chat(prompt)
}I will run this twice with no seed value.
noseed <- replicate(n = 2, {
generate_response(
system_message = "You are a helpful assistant.",
prompt = "Generate a short excerpt of news about a journey to Mars",
model = "gpt-3.5-turbo-1106",
seed = NULL
)
})[1] "LLM RESPONSE =========================="
"Earlier today, the Mars One mission successfully launched its first crewed spacecraft towards the red planet. The crew of six astronauts will
embark on a historic journey, with the goal of establishing a permanent human settlement on Mars. The mission, expected to take approximately
nine months, marks a significant milestone in space exploration and is sure to captivate the imagination of people around the world. Stay tuned
for updates on this exciting journey to the final frontier."
[1] "LLM RESPONSE =========================="
"NASA's Perseverance rover successfully landed on Mars today, marking a major milestone in the exploration of the red planet. The rover will
conduct crucial experiments and collect samples to search for signs of ancient life and pave the way for future human missions. This historic
journey to Mars offers hope for new discoveries and advancements in space exploration."Using Jaccaard similarity this has a similarity value of:
[1] 0.2588235Now we repeat with with a random seed value chosen:
withseed <- replicate(n = 2, {
generate_response(
system_message = "You are a helpful assistant.",
prompt = "Generate a short excerpt of news about a journey to Mars",
model = "gpt-3.5-turbo-1106",
seed = 12345
)
})[1] "LLM RESPONSE =========================="
"NASA's Perseverance rover successfully landed on the surface of Mars today, marking a significant milestone in the mission to search for signs
of ancient life on the red planet. The journey, which began seven months ago, culminated in a nail-biting descent and landing in the Jezero
Crater. Scientists and engineers are now eagerly awaiting the first images and data from the mission, which aims to pave the way for future human
exploration of Mars."
[1] "LLM RESPONSE =========================="
"NASA's Perseverance rover successfully landed on the surface of Mars today, marking a significant milestone in the mission to search for signs
of ancient life on the red planet. The journey, which began seven months ago, culminated in a nail-biting descent and landing in the Jezero
Crater. Scientists and engineers are now eagerly awaiting the first images and data from the mission, which aims to pave the way for future human
exploration of Mars."And we get a Jaccard Index of:
[1] 1This minor step has allowed me to rerun the workflow and achieve the same response. Making these workflow changes I think will go a long way to making GenAI outputs more transperent and allow robust discussions around how they work and how they can be used to help decision making.
Note: A more rigourous test with better similirty measurement (using embeddings) of this is done in this post: https://cookbook.openai.com/examples/reproducible_outputs_with_the_seed_parameter
Looking for a data science consultant? Feel free to get in touch here