Reinforcement Fine-Tuning—12 Days of OpenAI: Day 2

hi everyone my name is Mark and I lead research at openai yesterday we took 01 out of preview and we launched it in chbt we're soon going to launch it in the API if you haven't been following o1 it's our latest series of model improvements that allow the models to think for a while before they come back with a response today we're really excited to preview our latest advancement in our model customization program it'll let users find to 01 on their own data sets and again this isn't standard fine tuning this is reinforcement fine tuning which really leverages the reinforcement learning algorithms that took us from Advanced High School level to expert PhD level for your own use cases I want to stress again that this is a preview of something that we're going to launch publicly next year but if you are a university or you're a researcher or you're an Enterprise we'll give you some information on how you can access our output program later so why would you want this thing well it allows you to take your golden data sets and turn them into unique offerings that will give you the same magic that we have for your own users and your own customers so I'll let John Julie and Justin say a little bit more yeah hello everyone yeah my name is John Allard and I'm an engineer here at openi hi everyone I'm Julie W I'm a researcher here at open AI I'm Justin ree I'm a computational biologist at Berkeley lb today we're so excited to be introducing this new way of model customization for our 01 series of models uh reinforcement fine tuning or rft for short for the first time developers researchers and machine learning Engineers will be able to use reinforcement learning to create expert models capable of excelling at their specific tasks within their domain we believe that any field which requires deep expertise in their AI models stands the benefit so if you work in say legal Finance engineering Insurance uh this one's for you for example we recently partnered with Thompson Reuters to use reinforcement fine-tuning to fine-tune 01 mini to be a legal assistant in their co-counsel AI um this tool assist their legal Professionals in accomplishing some of their most analytical workflows yeah so some of you will be familiar with the supervised fine tuning API that we launched um early last year and supervised fine tuning is really powerful what you're trying to do is get the model to replicate features that it finds in uh input text or images and this is great if you want to change the tone or the style or the response format of the model now with reinforcement fine-tuning or reindeer enforcement fine tuning I should say the with reinforcement fine tuning um it's actually it's different so you're not just teaching the model to mimic its um its inputs what you're teaching it to do is to learn to reason in entirely new ways over custom domains and the way this works is that we when the model sees a problem we give it space to Think Through the problem and then we grade the final answer from the model and then using the power of reinforcement learning we reinforce lines of thinking that led to correct answers and we disincentivize incorrect answers um and what you'll see is that you know with as little as a few dozen examples the model will learn to reason in new and effective ways over custom domains that's crazy that you can do that with just 12 examples that's not something you can do with regular uh fine tuning yeah exactly yeah in the space of large language models and large machine learning a few dozen examples is basically nothing yeah so for the first time our model customization platform will support reinforcement learning and notably this is the same technique that we use internally at open AI to train our Frontier models like gp40 and the o1 series one area with many exciting applications is scientific research but don't just take our word for it that's why we're joined today by Justin ree uh Justin is a researcher at Berkeley lab and one of his areas of study is using computational methods to understand the Gen the genetic causes underlying rare diseases Justin thank you so much for being here do you mind telling us a little bit more about your research and how reinforcement fine tuning might help sure thanks it's great to be here so one of the areas of my research is rare genetic disease so contrary to the name rare genetic disease is actually not rare So any one rare disease is rare but if you put them all together um they're actually quite common and so we're talking about 300 million people globally who suffer from a rare disease and what's more these people often have a long diagnostic Odyssey of months to years before they find out about their condition W it's like uh the whole population of the US yes it's not a small number of people and so what we're working on is better uh computational tools and methods to Really research what's important uh and to help us understand and treat these diseases so we do our work in kind of an academic setting and learning more about the rare disease and their causes and the hope is we'll be able to advance the healthcare for these folks uh going down the line and now assessing rare disease

is kind of hard because you kind of have to have two things you have to have sort of expert domain knowledge about the medical side of things and you also have to have uh sort of systematic reasoning over the biomedical data and this is an area where we think that o the oan model can really help us out with its reasoning capabilities that makes a lot of sense you know our large language models have domain knowledge and our o1 models are really systemic reasoners so it seems like now there's a pretty good computational method for addressing some of these that's right can you tell us a little bit more about the data sets that you're using sure so this was sort of a collaborative effort between uh our group and charot Hospital in Germany and Peter Robinson's lab and the Monarch initiative um and what we did really was to extract disease information from hundreds of scientific Publications that were case reports about rare disease and so uh we sort of curated the information uh and that's lists of signs and symptoms that were present in the patient and that were excluded then of course the disease that they had and importantly for the conversation the causative Gene that was mutated that was causing the problems in these fols I see so you and maybe some doctors are trying to figure out given a patient symptoms uh What gene might have mutated to cause those symptoms yeah that's right and so something we've been working on together with the open AI team is uh training the old one models to reason more effectively about the causes of disease incredible thank you Justin uh we're now going to give you a preview of reinforcement fine-tuning at work and not to steal any Thunder but we're going to take 01 mini and make it exceed the performance of 01 on this task uh that's the 01 that we just launched yesterday and this matters so much because o1 mini is a smaller faster and cheaper model than 01 yeah so using Justin's data set we're going to show you can just drastically improve the performance of ow and mini um on this task where given a list of symptoms you're trying to predict which Gene might be responsible for the genetic disease and so to give an overview of this process we're going to start by looking at uh data sets that are used to train the model and graders evaluate the model and then we're going to um launch a training job on open AI training infrastructure and finally we'll evaluate the resulting fine-tune model so we can see how it's improved over the base model that we started with so to start us off we're going to jump over to the open AI development platform and we're going to go ahead create a new model so um you know we've had supervised fine tuning for a bit over a year now what we're going to do is we're going to select reinforcement fine tuning now we're going to be training o1 so we'll select that as the base model and now we need to upload a training data set and now training data sets they're just Json L files which is just a file where each line in the file is an example that you want the model to be trained on for this um case Justin and his colleagues assembled a data set of about 1100 examples um and so I'll go ahead and upload that one and just so that we get a really good feel for um how this data set works and what this task is we'll zoom in on an individual data point really quickly and so this is what an individual data point looks like and there's really three important things here so the first is the case report and this is a description of the patient and the patient symptoms so we see that the patient was a 51-year-old woman the disease onset was not specified we have a list of symptoms like hyperism um hyperthyroidism and others um as Justin said earlier we have the absent symptoms um these are the symptoms that are not present and this is important because it helps the model to rule out genes that it might think would otherwise be responsible um for the symptoms that are present next we have the instructions and I'm sure if you're watching this live stream you're familiar with prompting and so all we're doing here is just prompting the model for um what we wanted to do for this task and so what we're saying is you know given the list of symptoms and the case report can you list all the genes that you think might be responsible for the um for the genetic disease that you think is present and then we also asked it to provide an explanation for why it thinks those genes might be responsible um finally we also have the correct answer and so this is the gene that we happen to know is responsible but importantly we're not showing this to the model during the training process that would be cheating but we're using it internally during the training process to grade the model's outputs or to check if the model is correct this is a pretty hard task uh I definitely have no hope of answering this question yeah I mean you can tell that we're we've come up far away from just trying to count the number of RS in the word strawberry yeah so um now when we give the model this prompt this case report and these instructions the model is going to output something like this which is a list of genes that it thinks might be responsible and importantly the genes are in sorted order where the first Gene in the list is the one that it thinks is most likely to be responsible the second one in the list is the one that it thinks is second most likely and so on and so forth cool so um we'll hop back over and so um next we need to upload some validation data and validation data um it's going to be in the exact same format as the training data but importantly there's no overlap in the

correct genes between the validation data set and the training data set and what that means is that the model can't cheat it has to um or it can't learn to just memorize a list of symptoms and Associate those with the gene it has to actually generalize from the training data set to the validation data set gotcha so I mean where's the reinforcement part come in you know we talked about grading uh is that part of the process here yeah that's a really good question so grading is done by this concept of graders that we're introducing here and so um graders are really simple what a greater does is it takes the output from the model and it takes the correct answer and it Compares them and it returns a score between zero and one and so zero means that the model did not get the answer correct at all and one means the model got the answer correct and you can also give partial credit so it can be anywhere in that range so for this specific task we have a grader that looks like this so it takes the correct answer um that we happen to know and it takes the output from the model which is the list of genes and it produces a score so in this case you know foxy3 is the correct answer it was second in the list of genes and so it gets a score of like 7 I see so if it had instead said foxy 3 was first in the list I would have gotten a grade of one yeah exactly and then as it gets further and further along the list the score kind of gradually decays to zero ah nice makes sense um but what if I have a task that isn't you know grading A ranked list do we have other graders that are more General yeah so we're supplying kind of a collection of graders that we think pretty effectively cover the space of possible intents that you might have um you for while doing enforcement fine tuning and we're always adding more yeah and eventually we're going to hopefully let you define your own graders yeah maybe like upload a python file or something and do some custom grading yeah cool so um we've defined our training data set we've defined our validation data set let me go ahead and copy in the grader really quick um and now openi allows you to set um you know we allow you to customize these fine tuning runs by setting hyper parameters but we set some pretty good default so I'm just going to go ahead and click create here now what this is doing is um you know we we've just kicked off a training job um and so the really cool thing is that um you bring the data set and you bring the greater and these are the places where you really have domain expertise and where you can really contribute to this problem and then you get to leverage the full power of open AI reinforcement learning algorithms and our full distributed model training stack to customize a Frontier Model for your use case so as a user I just like to bring my data set and grader and open takes care of everything else yeah exactly yeah um so you know reinforcement F tuning jobs can take anywhere from a few hours to a few days to run so we're going to jump over to a job that I ran earlier this week on the same data set um just so we can kind of see the results so I'll jump over here um so I have this job that I ran earlier this week um it completed successfully It produced a fine two model for us and there's one thing that I want to um look at which is um the validation reward score and so what this is the it's the average score from the greater on the validation data set and how it changed over the course of the fine-tuning run and so what we can see is that the score is going up and as we said earlier since there's no overlap in genes between the training data set and the validation data set it means that the model really learned to generalize on our task um it wasn't simply memorizing a list of symptoms and mapping those to genes so you know while this is cool you know the chart goes up and to the right which is what we like to see um it'd be nice if we could get a better feel for how the model has actually changed during the fine-tuning process and so you know we'll take a closer look at that now all right so we're going to pop over to the evaluations dashboard which is a product in our developer platform that we launched earlier this year um there's a lot of numbers but don't worry we're going to go through all of them so I've set up three different runs here the first one was uh a run against our 01 model which we released yesterday the second was against 01 mini which was the starting point of our fine-tuning job and then finally uh the reinforcement fine tuned 01 mini now we looked at the reward going up and to the right but what does that actually mean for this task I've set up three different evaluations to sort of assess that the first one is top at one which is how often is the correct answer the very first item in the list top at five which is how often is the correct answer in the top five elements of the list and finally top at Max did we at all put the right answer in our list um and so looking at top at one we can see that our starting point ow and mini got 177% on our data set of about 200 01 got 25% so it's doing better uh but then our fine-tuned 01 mini got 31% awesome uh I took a screenshot of this and I put it into chat GPT and asked it to make me a plot a Christmas them plot and uh here's a nice visualization of those nine numbers that we saw earlier so you can see uh our starting point 01 mini uh across top at one top at 5 and top at Max our 01 model and then finally our best performing model which is this 01 mini fine tune

here in um dotted red line so looking at these results what do you think Justin well I think this is pretty impressive performance and especially the increase in the validation uh data because that implies that the model is learning something generally about how to reason over these kind of data which is pretty exciting um and so an obvious question you might ask is how is this doing comparing compared to existing bioinformatics tools and I don't really have an Apples to Apples comparison but um because typically in this kind of experiment you would provide uh genomic sequencing data and we haven't included that here um but the sort of open-ended quering of models here over incomplete symptom list is new and exciting I think great uh so these are aggregate statistics but let's look at the actual model responses so I'm going to pop on over to this data tab let's filter by the passes uh and so here is the input that we're giving to the model so the problem as John described earlier is to identify genes that may be responsible for a set of observed symptoms uh we asked the model to Output a dictionary containing both a string that explains you know why did I pick these genes and of course the genes themselves in ranked order and then finally we have the symptom list as well so this um patient presented with sub endal nodules seizure uh yeah and a couple other things uh we then run our models so this was our 01 model this one our fine-tuned o1 Mini model um we gave it that input and now the output is uh this dictionary that we described earlier so reasoning the combination of sub endal nodules uh seizure cortical tubulars are indicative of this complex which is commonly caused by mutations in these genes um it lists a couple other potential ones and then it says tsc2 is the most likely um candidate and if we scroll back on over to our answer we'll see that tsc2 is in fact the correct answer so that allowed us to get a pass on top at one top at five and top at Max so looking at this output um Justin like is this a useful output for the model to be giving back yeah absolutely so it's particularly useful to see the model's reasoning and that's a big contribution here and also the r obviously the rank list of answers so even if the the correct answer is not first you know you can look at all the possibilities um and it's also great to see the fine-tuning improves the performance of in the rank list of possible answers so that right answer is getting closer to one so that's gratifying Justin kind of zooming out a little bit like how does reinforcement learning shape your field um can you talk about some Trends in viy sure so I think there's a lot of interest in the research community and Import in using these models for these kind of tasks and so the feeling for this particular use case is that the best solution in the near term is probably a hybrid solution between existing bioinformatic tools and these uh models like 01 uh and so this is I think is excellent progressing sort of characterizing the strengths of these models and and also how we can use uh tools like fine-tuning to improve performance um and so like I said there's not really a comparable Benchmark to compare the two but um it it's definitely progress and how we can use these models to kind of understand disease and then you know in a larger sense to sort of how we can incorporate these models into a workflow that will eventually improve healthcare for these folks right amazing thank you Justin so while we've just shown you an exciting application of reinforcement fine-tuning in scientific research this is a general purpose technique we've seen promising results in data sets from biocham from AI safety from legal um and from Health Care as well we can think of hundreds of more examples or tasks that we can use this model on but we know that you can probably think of many more uh so that's why we're so excited to be expanding uh our Alpha program today to enable more people to push the boundaries of the capabilities of our 01 models on the tasks that matter the most to them yeah so you know we've been working with a small group of trusted Partners to really test out reinforcement fine tuning and today we're expanding Alpha access via what we're calling the reinforcement fine-tuning research program so um you know this program is ideal for organizations who are currently working on very complex tasks with teams of experts and who think that they might benefit from AI assistant on these tasks so you know if you're interested in applying for one of these limited spots you can find a link to the application in the description of this live stream and as Mark said ear um earlier you know we plan on launching this product reinforcement fine tuning publicly early next year yeah we're all really truly excited to see what you do with reinforcement fine tuning and really speaking as a researcher there's nothing that makes us happier than seeing our models being adapted and used to advance you know science knowledge in the real world do you have a joke for us today well as so happens I do uh as it's

become a tradition I have a Christmas theme joke so you know we live in San Francisco self-driving vehicles are all the rage and actually Santa's been trying to get in on this too um he's trying to make a self-driving sleigh but he for some reason his models just keep not identifying trees and the Slate is hitting trees left and right uh do you guys have any guesses why no he didn't Pine tun his oh jeez okay all right um please join us next week we'll have a lot more to share thank you