Use a pre-completion document, that stores the cached response of the model to write to as a sort of reverse RAG (gives a space for multi-hop to happen, and stores context for cross document RAGs)
@FunCodingwithRahul Жыл бұрын
Can I integrate any LLM as generator model e.g. Falcon? I am getting errors while doing it. any lead would be of great help !!
@cosmicray007 Жыл бұрын
Question, why not just use Apache Hive as a Catalog meta database, like the one found in AWS Glue. No need to move the data, since the catalog mdb points to the data. Moving data is costly, just saying. 18:44 sounds like we are in a pickle :)
@johnryan8645 Жыл бұрын
Clifford algebras do high dimensional cross products.
@alexiagordon6389 Жыл бұрын
Wow i love this! You need Promo-SM!!!
@goelnikhils Жыл бұрын
Amazing Talk
@anupamaappasahebdeshmukh4395 Жыл бұрын
Thank you sir, do you also have some gihub samples to understand this concept? Can you please share?
@eugeneware3296 Жыл бұрын
This is so brilliant. It's so hard to find these real-life learnings of how to deliver world-class ML systems in production. Especially the interaction between spark and transformers.
@jasonmackay891 Жыл бұрын
100%
@VaclavKosar Жыл бұрын
Great meetup, thx. Loved the part about Spark cpu transformer embedding calculation for semantic indexing.
@mprone Жыл бұрын
Any chance to get the first github repository ? after more than 2 years it is still private I guess
@dibdias12 жыл бұрын
Very good to have this video!
@hritikakolkar2 жыл бұрын
Informative
@ppujari2 жыл бұрын
segmentation part is not clear. Lookslike speaker is not intended to tell.
@justingidman16842 жыл бұрын
ᴘʀᴏᴍᴏsᴍ
@KwanyuetHo2 жыл бұрын
I think I still have the questions about merging cross products. From what I understand cross products between two vectors are only defined for three-dimensional vectors. Is there an extension to n dimensions? Is it something from differential geometry? Or are you talking about outer products?
@vikankshnath80682 жыл бұрын
What about multi span setting? When ans is spanned at different places.
@davidnkanta96652 жыл бұрын
Can i deploy my model as a Conversational Agent(like gpt3 or bert for question answering) on a mobile application. Further more, can i get graphical plot of the training and accuracy of my trained model??
@davidnkanta96652 жыл бұрын
For 6 montha i have been struggling with my dissertation project. This is pure magic
@doyourealise2 жыл бұрын
i could not attend the live session so i am here. Thanks for the video.
@doyourealise2 жыл бұрын
thaanks for the video
@giannagiavelli50982 жыл бұрын
Transformer is rubbish, years behind Noonean
@ajitkumar152 жыл бұрын
Thank you for such video which give us glance in the advancement of Language processing
@_HarshVerma2 жыл бұрын
Now this guy knows what he is talking about otherwise every other person talkin rag on KZread is just saying gibrish
@bhaumikpatel4902 Жыл бұрын
Probably because he was a co-author 😅
@ZKYQUQ6 ай бұрын
You're right bro. Love u.@@bhaumikpatel4902
@in_experience63833 жыл бұрын
Thanks man it was nice Intro. the Q&A beyond SQuAD.
@digidim3 жыл бұрын
The repository does not exist. Could you please recheck the github link ?
@asitkumar31763 жыл бұрын
why dont u post the code about which you are talking about..I mean with retrievers
@DistortedV123 жыл бұрын
Awesome work
@jamesheffernan80003 жыл бұрын
Excellent, Thank You Dave.
@connorshorten63113 жыл бұрын
These videos are great! Thank you so much!
@gulanshunan3 жыл бұрын
what if an answer is not in the given text
@paraschopra3 жыл бұрын
You use abstractive answer generation technique.
@gulanshunan3 жыл бұрын
what to do if an answer is not in the given text
@gulanshunan3 жыл бұрын
will you share pdf version or repository of code
@TheMriganks3 жыл бұрын
Good session, possible to get the slides?
@venkatagudala51724 жыл бұрын
51:53 Retriever in the QA Pipeline uses modified TF-IDF which is Okapi-BM25 by adjusting the Term Frequency(TF) and Inverse Document Frequency (IDF) by adding parameters 'k1' and 'b' to the TF-IDF equation
@connectrRomania4 жыл бұрын
amazing, is there a pdf or ppt version of the presentation?
@katejannuzzi23754 жыл бұрын
Thank you for posting this to you tube, I was happy to see this after I missed the live one!
@gabrielaltay58074 жыл бұрын
Hello viewers, There was a window sharing issue between 24:30 and 27:30. The kernel that was supposed to be on the screen was an introduction to the Wikipedia part of the dataset (www.kaggle.com/kenshoresearch/kdwd-wikipedia-introduction). Also we didn't have time to get to many of the kernels, but you can check these out ... * www.kaggle.com/kenshoresearch/kdwd-explicit-topic-models * www.kaggle.com/kenshoresearch/kdwd-pmi-word-vectors * www.kaggle.com/kenshoresearch/kdwd-aliases-and-disambiguation * www.kaggle.com/gabrielaltay/kdwd-subclass-path-ner and a blog post going into more detail on creating synonyms from Wikipedia blog.kensho.com/how-to-build-a-smart-synonyms-model-1d525971a4ee
Пікірлер
Use a pre-completion document, that stores the cached response of the model to write to as a sort of reverse RAG (gives a space for multi-hop to happen, and stores context for cross document RAGs)
Can I integrate any LLM as generator model e.g. Falcon? I am getting errors while doing it. any lead would be of great help !!
Question, why not just use Apache Hive as a Catalog meta database, like the one found in AWS Glue. No need to move the data, since the catalog mdb points to the data. Moving data is costly, just saying. 18:44 sounds like we are in a pickle :)
Clifford algebras do high dimensional cross products.
Wow i love this! You need Promo-SM!!!
Amazing Talk
Thank you sir, do you also have some gihub samples to understand this concept? Can you please share?
This is so brilliant. It's so hard to find these real-life learnings of how to deliver world-class ML systems in production. Especially the interaction between spark and transformers.
100%
Great meetup, thx. Loved the part about Spark cpu transformer embedding calculation for semantic indexing.
Any chance to get the first github repository ? after more than 2 years it is still private I guess
Very good to have this video!
Informative
segmentation part is not clear. Lookslike speaker is not intended to tell.
ᴘʀᴏᴍᴏsᴍ
I think I still have the questions about merging cross products. From what I understand cross products between two vectors are only defined for three-dimensional vectors. Is there an extension to n dimensions? Is it something from differential geometry? Or are you talking about outer products?
What about multi span setting? When ans is spanned at different places.
Can i deploy my model as a Conversational Agent(like gpt3 or bert for question answering) on a mobile application. Further more, can i get graphical plot of the training and accuracy of my trained model??
For 6 montha i have been struggling with my dissertation project. This is pure magic
i could not attend the live session so i am here. Thanks for the video.
thaanks for the video
Transformer is rubbish, years behind Noonean
Thank you for such video which give us glance in the advancement of Language processing
Now this guy knows what he is talking about otherwise every other person talkin rag on KZread is just saying gibrish
Probably because he was a co-author 😅
You're right bro. Love u.@@bhaumikpatel4902
Thanks man it was nice Intro. the Q&A beyond SQuAD.
The repository does not exist. Could you please recheck the github link ?
why dont u post the code about which you are talking about..I mean with retrievers
Awesome work
Excellent, Thank You Dave.
These videos are great! Thank you so much!
what if an answer is not in the given text
You use abstractive answer generation technique.
what to do if an answer is not in the given text
will you share pdf version or repository of code
Good session, possible to get the slides?
51:53 Retriever in the QA Pipeline uses modified TF-IDF which is Okapi-BM25 by adjusting the Term Frequency(TF) and Inverse Document Frequency (IDF) by adding parameters 'k1' and 'b' to the TF-IDF equation
amazing, is there a pdf or ppt version of the presentation?
Thank you for posting this to you tube, I was happy to see this after I missed the live one!
Hello viewers, There was a window sharing issue between 24:30 and 27:30. The kernel that was supposed to be on the screen was an introduction to the Wikipedia part of the dataset (www.kaggle.com/kenshoresearch/kdwd-wikipedia-introduction). Also we didn't have time to get to many of the kernels, but you can check these out ... * www.kaggle.com/kenshoresearch/kdwd-explicit-topic-models * www.kaggle.com/kenshoresearch/kdwd-pmi-word-vectors * www.kaggle.com/kenshoresearch/kdwd-aliases-and-disambiguation * www.kaggle.com/gabrielaltay/kdwd-subclass-path-ner and a blog post going into more detail on creating synonyms from Wikipedia blog.kensho.com/how-to-build-a-smart-synonyms-model-1d525971a4ee