LLMs have low randomness: if you ask the same thing twice you get similar responses. Generator prompts are a way to boost the randomness of LLMs.
Using a few generator prompts, I had Gemini write an entire instruction tuning dataset from scratch. It outperform popular datasets.
Let’s start with a toy example of why we need generator prompts. Suppose I want a list of different colors. So I feed this prompt to Gemini 1000 times. This does poorly - I only get 33 unique outputs from 1000 runs. I need more randomness.
A generator prompt asks the model to enumerate a long list of execution paths, and then randomizes which paths get chosen.
Here's an example. The numbers 23 and 76 are randomized each time the prompt is called.
This prompt gives me 782 unique outputs from 1000 runs.
Now let’s do something useful. I want to extract math questions/answers from Gemini. I’ll ask the model to enumerate a long list of different topics, then subtopics, then write me a question.
Run this prompt through Gemini Pro 500K times, and presto! The resulting math instructions out-perform the human-curated MathInstruct dataset.
The GenQA dataset contains 11M questions in 9 splits. You can find it on HuggingFace.
Finally, shout out to the recent Magpie dataset. It creates instructions by giving an empty prompt to Llama. Our “general” split was created the same way, but using GPT-3.5 instead of Llama. We focused on generator prompts because they are controllable, but empty prompts produce great "general" instruction sets like Magpie.
There's also a related paper from Yiming Zhang, @A_v_i__S, Nicolas Carlini, @zicokolter, and @daphneipp on prompting strategies for helping LLMs make random choices.
LLMs have low randomness: if you ask the same thing twice you get similar responses. Generator prompts are a way to boost the randomness of LLMs.
Using a few generator prompts, I had Gemini write an entire instruction tuning dataset from scratch. It outperform popular datasets.Let’s start with a toy example of why we need generator prompts. Suppose I want a list of different colors. So I feed this prompt to Gemini 1000 times. This does poorly - I only get 33 unique outputs from 1000 runs. I need more randomness.A generator prompt asks the model to enumerate a long list of execution paths, and then randomizes which paths get chosen.
Here's an example. The numbers 23 and 76 are randomized each time the prompt is called.
This prompt gives me 782 unique outputs from 1000 runs.Now let’s do something useful. I want to extract math questions/answers from Gemini. I’ll ask the model to enumerate a long list of different topics, then subtopics, then write me a question.Run this prompt through Gemini Pro 500K times, and presto! The resulting math instructions out-perform the human-curated MathInstruct dataset.The GenQA dataset contains 11M questions in 9 splits. You can find it on HuggingFace....or you can read our tech report on how to quickly distill your own bespoke instruction datasets using powerful models.Finally, shout out to the recent Magpie dataset. It creates instructions by giving an empty prompt to Llama. Our “general” split was created the same way, but using GPT-3.5 instead of Llama. We focused on generator prompts because they are controllable, but empty prompts produce great "general" instruction sets like Magpie.There's also a related paper from Yiming Zhang, @A_v_i__S, Nicolas Carlini, @zicokolter, and @daphneipp on prompting strategies for helping LLMs make random choices.
yes
LLMs have low randomness: if you ask the same thing twice you get similar responses. Generator prompts are a way to boost the randomness of LLMs.
Using a few generator prompts, I had Gemini write an entire instruction tuning dataset from scratch. It outperform popular datasets. ... Let’s start with a toy example of why we need generator prompts. Suppose I want a list of different colors. So I feed this prompt to Gemini 1000 times. This does poorly - I only get 33 unique outputs from 1000 runs. I need more randomness. ... A generator prompt asks the model to enumerate a long list of execution paths, and then randomizes which paths get chosen.
Here's an example. The numbers 23 and 76 are randomized each time the prompt is called.
This prompt gives me 782 unique outputs from 1000 runs. ... Now let’s do something useful. I want to extract math questions/answers from Gemini. I’ll ask the model to enumerate a long list of different topics, then subtopics, then write me a question. ... Run this prompt through Gemini Pro 500K times, and presto! The resulting math instructions out-perform the human-curated MathInstruct dataset. ... The GenQA dataset contains 11M questions in 9 splits. You can find it on HuggingFace. ... ...or you can read our tech report on how to quickly distill your own bespoke instruction datasets using powerful models. ... Finally, shout out to the recent Magpie dataset. It creates instructions by giving an empty prompt to Llama. Our “general” split was created the same way, but using GPT-3.5 instead of Llama. We focused on generator prompts because they are controllable, but empty prompts produce great "general" instruction sets like Magpie. ... There's also a related paper from Yiming Zhang, @A_v_i__S, Nicolas Carlini, @zicokolter, and @daphneipp on prompting strategies for helping LLMs make random choices.
Missing some Tweet in this thread? You can try to
Update
More Threads by @tomgoldsteincs
Diffusion models are powerful tools for image creation, working by adding noise to images and then removing it to genera...
