-
Notifications
You must be signed in to change notification settings - Fork 3.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Creating augmented data using few-shot prompts for explanations of jokes, logical inferences, etc. #261
Comments
Adding expalantions at the end of existing instruction dataset answers where the answers are classificaitons (see p3, natural instructions, etc): For exmple, This is a movie review for the movie {movie}: {review}. This movie review is {classifciaiton} because ...[your created answer] This is a movie review for the movie {movie}: {review}. This movie review is {classifciaiton} because ...[your created answer] This is a movie review for the movie {movie}: {review}. This movie review is {classifciaiton} because ...generated answer We can also follow this up with explanations for other "hard" things like: explain riddles, poems (metaphors), analogies, songs |
Going with the movie reviews idea, could we use the Rotten Tomatoes dataset to generate prompts, maybe supplement with one of the models fine tuned on it as well? |
The idea is to create a dataset with explanations. Like for example take the movie dataset and do this: |
@momegas yes. if it is very compute intensive, it doesn't need to be large. maybe see if you can get it to work first. And then we can discuss size. we can run it on some extra compute. |
Sound like a very cool task and I would love to give it a try if it is still relevant :) @ontocord |
@ontocord I'd like to have a try, can you tell me your name in Discord? Maybe we can talk a little bit more there. |
Regarding #261. This is an re-produce of the dataset from LogicInference Dataset in paper: https://openreview.net/pdf?id=HAGeIS_Lcg9. I think it will helpful for improving logic inference ability of the model. The github page of LogicInference Dataset: https://github.com/google-research/google-research/tree/master/logic_inference_dataset. This dataset is aimed to offer more dataset for Open Assistant project, depending on their demands, there three columns: INSTRUCTION, RESPONSE, SOURCE. The results in this dataset is a little different from which was introduced in the original paper: 1.For all three splits (IID/OOD/length), only IID is used. In the original paper, it seems that model can reach better performance with data generated by this split method. 2.In the original paper, there are two form of responses: LOGICINFERENCEb (with the answer at the beginning) and LOGICINFERENCEe (with the answer at the end). This dataset uses LOGICINFERENCEe, that means: for all questions, the model will first do logic inference, and give the final answer at the end. 3.The original paper, some parameters in generate_dataset.py are: N_INFERENCE_PROBLEMS = 5000 N_VARIATIONS = 25 N_EXAMPLES = 200000 TRAIN_RATIO = 0.9 LENGTH_SPLIT_THRESHOLD = 4 RANDOM_SEED = 0 I choose some new parameters: N_INFERENCE_PROBLEMS = 10000 N_VARIATIONS = 25 N_EXAMPLES = 55000 TRAIN_RATIO = 1 LENGTH_SPLIT_THRESHOLD = 4 RANDOM_SEED = 1111 The original script generated 4814 different inference problems and extended all those inference problems to around 200,000 Q-A pairs. My settings generated 5491 different inference problems and extended them to around 54,607 Instruction-Response pairs. I think for Open Assistant projects, maybe the number of different inference problems is more important, and generated many similar Instruction-Response pairs will only add training time and doesn't make much sense. --------- Co-authored-by: Andreas Koepf <andreas.koepf@provisio.com> Co-authored-by: Oliver Stanley <olivergestanley@gmail.com>
Going to work in this field, but with more specific tasks (semantics, logic, reasoning) |
Closing old data issue. |
See https://www.lesswrong.com/posts/EHbJ69JDs4suovpLw/testing-palm-prompts-on-gpt3.
Try doing 2, 3 or 4 shot inference on something like JT or neox 20B or galactica.
After we find a promising model and configuration, we can scrape the net for jokes and paragraphs with logical inferences to create dialog data.
Human: Tell me a joke about {extract keywords from joke}
Assistant: {joke}
Human: Explain the joke.
Assisant: {explanation}
See also https://storage.googleapis.com/pathways-language-model/PaLM-paper.pdf
The text was updated successfully, but these errors were encountered: