Nice! In my experience preference tuning with the ultra feedback datasets does not really change benchmark scores (and sometimes even makes them worse) but it does seem to improve the real-world user experience when chatting with the model.
I'm also not sure if orpo only on UF is better than sft on UC + DPO on UF, especially if you're also trying to do language adaptation. That, or first continue pretraining the model and then doing orpo.
While the "rules" of OpenAI do get frustrating from time to time, I do not blame others who do not follow the same path as I do. If I am asked why my licenses are different from someone else's I will answer according to what I've written in the post above (the rules suck and our vague, I understand why people do what they do and I do what I do because of other reasons). But I definitely do not want to go around and point fingers pre-emptively in hopes that people just use my models. Our community for Dutch is already quite small so I rather just lift each other up and build on each others work through friendly "competition" than to compete in bad faith.
So I think that for my future models, I'll just make use of ultrachat+ultrafeedback, which should be cleared for training apache 2.0 models because they were created with Azure. This may negatively impact the model's performance (especially for code because it does not include the Stack Overflow set) but I hope the impact is limited.
What do you mean with compliance in this context? I'm not sure how I can market being non-commercial as a good thing 😅
Cool! Looking forward to what you'll build with this!
I think the correct place for this is to make a new issue on their issue tracker: https://github.com/ScandEval/ScandEval/issues
scandeval) to run your own benchmarks with. As part of project "Leesplank" (with Michiel Buisman and Maarten Lens-FitzGerald) we recently added GPT-4-1106-preview scores to add a good "target" to the leaderboard.
Understandable. I'm especially attracted to the broad vocabulary, which can be of use for language adaptation.
What kind of weird results? In terms of loss, or really qualitative output?
Indeed, there is not a lot of metadata. There's also a discrepancy between the no. scores/languages and the no. paragraphs in the text. I've notified the authors about that. CulturaX is an attractive dataset, too!
load_dataset("BramVanroy/hplt_mono_v1_2", "nl_cleaned")Really cool, crowd-sourcing like this can be very powerful!
Just joined and I have one question/recommendation: the name "prompt" is ambiguous. I got the following conversation:
Marv is a chatbot that reluctantly answers questions with sarcastic responses:
You: How many pounds are in a kilogram?
Marv: This again? There are 2.2 pounds in a kilogram. Please make a note of this.
You: What does HTML stand for?
Marv: Was Google too busy? Hypertext Markup Language. The T is for try to ask better questions in the future.
You: When did the first airplane fly?
Marv: On December 17, 1903, Wilbur and Orville Wright made the first flights. I wish they’d come and take me away.
You: What is the meaning of life?
Marv: I’m not sure. I’ll ask my friend Google.
You: Why is the sky blue?
Intuitively, when I am asked to "rate the prompt" I would expect to have to rate the user prompt that is used to trigger a response, so does that mean I should rate "How many pounds are in a kilogram?" Or do I have to rate all the responses of "Marv"? Or do I have to rate the whole conversation? The guidelines are also not very strictly clear to me because the example that I get is a conversation, so there is a lot potential to be rated (one user prompt, all user prompts, whole conversation, etc.)
Hope you see my confusion. To make sure that everyone is rating the same aspects, this could be clarified!
From my limited experience, looking at potential unwanted text properties of chosen v. rejected can be another one to investigate. I think the model may learn, for instance, that longer sequences are better and will therefore learn to generate longer sequences regardless of the quality of the content. You can catch this in the metrics, I believe, but only in the log probs (which would then also be very low for the chosen text, perhaps even lower than the rejected text). You'll likely not notice this in the rewards metrics.
Update (since my post can't be too long): it seems that beta=0.5 was too high! Looking at the log probabilities, you'll notice that despite high rewards margins the log probabilities of all beta=0.5's are inconsistent (chosen log probs are lower than the rejected ones!). I am not sure what causes this. Perhaps the model has become good at discerning the two texts on other characteristics (like text length) but isn't really certain about the probabilities? In any case, I think that 0.1 beta is better. Perhaps there is some middle ground that I missed during grid search but looking at the results qualitatively, it seems that 0.1 is a good model.
/*****/). There were some useful brainstorms in that thread. I think the dataset is relatively easy for the model, leading it to quickly overfit when the beta is very small, which allows the model to step away further from its initially outputs.
That's also a good idea. I'm now going to have a look at the results of the different hyperparameter runs and then look into some example generations. I'll make a new post later where I share all the WANDB logs so everyone can have a look at the impact of lr, batch size and beta on all the losses and rewards. Currently it seems that a beta of 0.01 is just too low for how easy my dataset is.
@jvdgoltz @robinsmits I updated the link in the OP with a more permenant one, that should work!
I work on an already finetuned version of Mistral (GEITje) and on a Dutch dataset. Currently I am thinking that the dataset is "too easy" and that the differences between the chosen and rejected answers are simply to obvious. Currently doing a hyperparameter search similar to HF's blog post. I do see a lot of difference in logp between runsm going from -500 to -1100.
So I tweaked the learning rate to 1e-7 (was 5e-7), removed portion of the dataset (I used two datasets, one based on Orca Pairs, the other based on UltraFeedback, I kept the latter), and I decreased the max length to 2048 max_length 1152 prompt length. EDIT: I thought this had worked, but apparently I am wrong. For some prompts, the same result occurs with repetitions of /******/ everywhere. I am very confused about this and should find time to dig deeper but it is a tedious trial and error process of training and testing that east up a lot of my time. If anyone wants to have a look, I can provide gated access to the model.
Mostly cc @lewtun but also a shout out to everyone in this thread for brainstorming along!
my first name at argilla.io
Mail sent!
Sure, where should I reach you?
I think that this is a sensible point but looking at other DPO datasets (e.g. https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized) this should not be "required" for DPO to work. So before I jump into this rabbit hole I'm first looking into other options. Thanks for brainstorming!
Good point. I have looked at the data and set the lengths to more reasonable values wrt the dataset (ended up with 1536 for the prompt, 2048 max length for the responses). Having heteregoneous batches might indeed also be something to look into.
I have also looked at the dataset and in terms of quality I would expect it to be usable enough when I compare it to the UltraFeedback dataset.
So perhaps DPO is just very sensitive to hyperparameters. I'll try a number of different hyperparameters and see whether I can find some suitable ones. (With limit disk space this is not easy though, every checkpoint taking up almost 100GB!)
I'll report back with my findings.
Sadly it is not possible to rely on something like PairRM because I am not working on English. I am having a closer look at the dataset but it does not look that different from the (English) UltraFeedback in terms of quality (in my language) so it's hard to pinpoint the difference. I think I'll try out some different hyperparameters and see if I can find something stable.
I used the alignment-handbook hyperparameters for Zephyr (beta 0.01 for 1 epoch) and the architecture is also based on Mistral. The only change I made is set max_length=8192 because I didn't understand why it was set to 1024. But I don't know if length would impact the result so much?
/*****/. So despite the good performance on the DPO objective and strong scores on the validation set (no overfitting), something seems to go wrong. Perhaps the outputs are too different and the task is too easy, in which case DPO is not useful. But why then would the model start hallucinating and repeating the same token over and over again?
That's odd, the invite should be valid for 7 days. Let's try again: https://discord.gg/Zj7ptGnC
