Alvaro Bartolome
@alvarobartt
Followers
1,208
Following
485
Media
111
Statuses
1,101
☁️ cloud machine learning @huggingface and open source passionate
Barcelona, Spain
Joined August 2020
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
HAPPY TWENTY FOURTH
• 482603 Tweets
Chelsea
• 123112 Tweets
#ブルアカらいぶおーたむSP
• 122728 Tweets
Liverpool
• 108559 Tweets
God Kabir Increases Age
• 86846 Tweets
#光る君へ
• 82145 Tweets
BUILD TEAM FJR
• 50760 Tweets
ジャニーズ
• 48815 Tweets
ピーコさん
• 46079 Tweets
#海に眠るダイヤモンド
• 41703 Tweets
nayeon
• 40465 Tweets
Wolves
• 30336 Tweets
おすぎとピーコ
• 28345 Tweets
サクラコ
• 23696 Tweets
#LINEMUSICでキンプリ
• 22937 Tweets
リラックマ
• 22598 Tweets
#NHKスペシャル
• 20463 Tweets
オッドストーリー
• 19978 Tweets
CHOOSE TO BE KA-TROPA
• 19436 Tweets
日本シリーズ
• 18808 Tweets
Rangers
• 17619 Tweets
双子の弟
• 16728 Tweets
LMSY TAOBAO LIVE
• 16206 Tweets
日曜劇場
• 14110 Tweets
GenG
• 13574 Tweets
Flyquest
• 11579 Tweets
アイドルマリー
• 10972 Tweets
Pinned Tweet
🔥 ICYMI last week we created a comprehensive guide on deploying
@google
Gemma2 with multiple LoRA adapters using
@huggingface
Text Generation Inference (TGI) on Google Kubernetes Engine (GKE)!
- 💪 Serve multiple specialized models on a single GPU for maximum efficiency
- 🎯
0
2
8
🎉 2023 is not over yet... and we still managed to ship our last LLM of the year!
Notux 8x7b is a DPO fine-tune on top of Mixtral 8x7b Instruct (SFT + DPO) and it outperforms the rest of the MoE-based models on the 🤗 Hub
12
36
279
🪁 Zephyr 7B Beta is already available in
@Apple
MLX and hosted in the 🤗Hub!
It’s amazing how easy and blazingly fast is to convert the weights and run LLMs locally in MPS with minimal effort thanks to MLX
Kudos to
@awnihannun
and team!
2
27
188
Really excited about the release of `distilabel` and to hear everyone’s thoughts on it! 🔥
It is a Python framework to both generate and annotate datasets using LLMs and novel methods such as JudgeLM, UltraFeedback, Self Instruct, and more to come
3
22
130
💥 We just released Notus 7B v1!
A DPO fine-tune on top of Zephyr SFT, but with a slight modification in the binarization of the UltraFeedback dataset, resulting in an improvement on MT-Bench and LM Eval Harness benchmarks!
🤗:
More information below 🧵
4
27
115
If you are into DL and CV you should consider sponsoring
@wightmanr
because he is a great individual and what he does for the open-source community is simply awesome! 🔥
1
14
61
⚠️ From
@argilla_io
we just released a new version of our UltraFeedback binarized dataset, removing some contamination coming from some TruthfulQA prompts being used within the training set
2
10
51
💡 Remember the issues with UltraFeedback from
@OpenBMB
spotted during the past weeks by
@argilla_io
and
@allen_ai
?
Well, those have already been solved and merged into a single dataset by
@hamishivi
👇🏻 Find it in the
@huggingface
Hub
1
7
36
✨I've just published a community post in the
@huggingface
blog on how to generate suggestions for your datasets in
@argilla_io
using Llama 2 via Inference Endpoints!
Check it out at
0
9
35
🎉 `safejax` is already public!
`safejax` is a Python package to easily serialize JAX/Flax models using `safetensors` tensor storage format from
@huggingface
instead of `pickle`.
Check it out at !
1
5
34
💥 Notus 7B v1 is already quantized thanks to the awesome
@TheBlokeAI
!
You can find those in the 🤗 Hub at:
- GGUF
- AWQ
- GPTQ
0
6
29
✨ Don't miss the latest distilabel release! Now one can reproduce the synthetic data generation pipeline from DEITA! Also, one can use LLM engines such as
@ollama
and
@MistralAI
endpoints with minimal effort via distilabel!
🎉⚗️ distilabel 0.5.0 is out!
Packed with new features and enhancements to build and improve LLM fine-tuning datasets with the latest techniques.
Let's see what's inside this release!
🔥 Deita components for evolution and automatic data selection of instruction tuning data
🧵
2
13
51
1
8
24
I just released , which is a repository explaining how to monitor any ML model with
@weights_biases
, in this case a simple
@PyTorchLightnin
CNN 🎉
If you are willing to start with either PyTorch Lightning, wandb or both, you should check it out!
2
6
24
🚨 Attention everyone! has changed the internal API that they were using to retrieve the historical data with no prior notice, so investpy is not working right now... I'll work on a patch ASAP!
14
1
22
💥 I've just published a post on how to generate a helpfulness dataset, similarly to HelpSteer from
@nvidia
, but using AI Feedback (AIF) from GPT-4 via `distilabel` instead of via human annotators
Read more about it in the 🤗
@huggingface
Hub post!
0
7
23
🚀 Lately
@argilla_io
we've worked hard on providing the best and most compliant version of the UltraFeedback dataset for DPO fine-tuning!
So on, we've decided to create a
@huggingface
collection including some small notes on the process of curating it!
0
6
21
🚀
@AnthropicAI
Helpful and Harmless (HH) preference dataset has now an extended version!
And I took some time to format it in a more standardised way, following the chat formatting and adding both the `prompt` and `prompt_id` columns!
🤗Hub Dataset at
3
4
21
🎩 For anyone out here building with distilabel, make sure to include the new `synthetic` tag within the datasets you upload to the 🤗 Hub so that those are easier to browse!
Synthetic data is going to be massively important in 2024, so we have recently launched a new tag on the
@huggingface
Hub to facilitate the discovery and sharing of synthetic datasets. To add this tag to your dataset card metadata, simply include the `synthetic` tag.
3
21
115
0
3
17
💨 Notus 7B v1 was recently released by
@argilla_io
!
After curating UltraFeedback and following Zephyr's recipe, we were able to get better results on both AlpacaEval and LM Eval Harness while almost on par in MT-Bench
🤗 Read the post below
0
4
17
At
@argilla
we seek high-quality data and we’ll continue our efforts on LLMs, always following a data-first approach.
All the code and resources used for Notus are available at , feel free to follow our progress and experiments there!
More coming soon! 🔥
1
2
14
🎉 I've just released `investiny` which is a tiny version of `investpy` with a function to pull historical data from while we work on fixing `investpy`!
Tiny, fast, and simple to use.
Check it out at
2
0
11
We just launched a Chat UI instance running Notux 8x7B in
@huggingface
Spaces 👏🏻
Feel free to give it a try!
Yesterday, we released Notux 8x7b v1, a Mixtral 8x7B Instruct v0.1 fine-tune using a second iteration of DPO and our cleaned and binarized version of the UltraFeedback dataset.
Today, we release a new space to chat with it 🚀
0
20
71
0
1
12
@andrey_cheptsov
@argilla_io
@winglian
Indeed we rely on the alignment handbook from
@huggingface
but we adapted it to be able to fine-tune Mixtral, we'll upload the recipe soon to our repository!
2
3
10
🎩 There is also a channel in the
@huggingface
Discord were you can engage with us, ask questions about the annotation process, about
@argilla_io
, or anything you want!
Kudos to
@vanstriendaniel
and
@lunarflu1
!
Join the channel in the following link
🔥 Data is better together🔥
At
@argilla_io
&
@huggingface
, we believe in the collective intelligence of the OSS AI community
So we have partnered to let everyone contribute to AI datasets!
You can start contributing now to the first initiative:
🧵
2
19
102
0
3
10
🚨 I think that mentioning
@Investingcom
and retweeting/liking what I post regarding investpy can be useful for to notice the community and get back with an answer as soon as possible!
1
4
9
🤗 Friendly reminder to use Notux in
@huggingface
Spaces via Chat UI (for free and limited time!)
We (
@argilla_io
) deployed Notux (Mixtral 8x7B MoE fine-tune) in
@runpod_io
using TGI, and then exposing that via Chat UI
📌 Go try it out!
0
3
9
who’s on the same boat? 🛶
at
@argilla_io
we firmly believe so, and we’ll be doing our best to create high quality and open datasets with permissive licensing!
Get ready for 2024: The Year of Open Machine Learning Datasets! 🌟
Here's a mega thread for all the datasets I'm showcasing this year. 🧵
4
12
85
0
3
9
🧪 Recently,
@Apple
MLX released 0.6.0 and comes with 4-bit quantization support, and you can also find the 4-bit weights in the 🤗 Hub!
It implies faster inference, less resources required, and eventually some quality loss
1
0
9
Additionally, I also created a collection in the 🤗 Hub with some 7B LLMs prepared to be used with MLX
0
1
9
TIL that you can see the requests sent to the Open LLM Leaderboard from
@huggingface
H4 at , and also check the status of yours (useful if it failed), h/t
@ClementDelangue
0
1
9
@pentagram
's work on
@cohere
type-face is awesome! the whole cohere blog-site and content is the most pleasing to read and go around by far, congrats! 👨🏻🍳
1
0
1
Just a reminder to everyone that may have missed it, I've released `investiny` which is a temporary replacement for `investpy` fully working.
Check it at
1
0
9
a lot of people asking about the monitoring! for those of you asking it’s `asitop` from
@timothy_lkh_
👏🏻
feel free to star their repository on github!
also kudos to
@awnihannun
as I discovered because of him!
0
0
8
👀 We're on almost 200 contributors already!
From
@argilla_io
we're extremely grateful for everyone that has contributed with the collaborative annotation so far, and would like everyone interested to join
And the data will be released with a permissive license (MIT) which is
🔥 Data is better together🔥
At
@argilla_io
&
@huggingface
, we believe in the collective intelligence of the OSS AI community
So we have partnered to let everyone contribute to AI datasets!
You can start contributing now to the first initiative:
🧵
2
19
102
0
3
7
Along with Zephyr they also released the code and recipes to easily reproduce Zephyr's experiments and also an easy way to run your own, and it's been incredibly helpful for researchers and developers alike.
Kudos here to
@_lewtun
and
@edwardbeeching
1
1
8
We would also like to thank
@OpenBMB
for generating UltraFeedback, which has been a really useful AI Feedback (AIF) preference dataset towards most of the alignment efforts out there!
1
0
8
@ptrblck_de
@NVIDIAAI
@soumithchintala
@PyTorch
Completely deserved! 🙌No ML practitioner that has worked with
@PyTorch
can say that you didn't help her/him out of a problem
0
0
7
But we were not the only ones shipping late 2023 stuff, as
@huggingface
launched the beta of the Hub Posts which are simply awesome, you should check those out!
1
2
8
🔊 As everyone may already know has included Cloudflare protection to their APIs, so investpy is no longer working... I've already tried to contact
@Investingcom
privately so as to see whether we can get to an agreement! I'll keep you all updated.
🚨 Attention everyone! has changed the internal API that they were using to retrieve the historical data with no prior notice, so investpy is not working right now... I'll work on a patch ASAP!
14
1
22
1
3
7
🤖 Hilo de distintos recursos y recomendaciones personales sobre IA 🤖
1
2
7
🤔 Do you struggle to serialize your JAX (Flax, Haiku, or Objax) model params?
`safejax` solves that by providing a unified way of serializing JAX model params using `safetensors` from
@huggingface
as the tensor storage format, instead of `pickle`.
2
3
7
👀 Also worth mentioning that anyone can track the annotation progress of the community!
But don't just focus on the numbers, focus on great and high quality annotations, we're sure you all will 💪🏻
🤗 Space at
🔥 Data is better together🔥
At
@argilla_io
&
@huggingface
, we believe in the collective intelligence of the OSS AI community
So we have partnered to let everyone contribute to AI datasets!
You can start contributing now to the first initiative:
🧵
2
19
102
1
4
7
investpy v1.0.7 is already out! 🔥
Make sure to update it using `pip install investpy==1.0.7` and let me know whether you like the new features or what do you want for the next release!
Go check it out and read the full release notes at
1
1
6
🧪 running some more experiments with notus, stay tuned for updates!
0
0
7
📚 Find more examples and useful resources in the `mlx-examples` collection of examples that both the MLX developers and the community are putting together
1
0
6
As an starting point we focused on what we do best at
@argilla_io
, seeking high-quality data, and realised that UltraFeedback had an issue in the overall score so we used the mean of the preference ratings instead.
🤗:
1
1
6
1
0
6
Then we fine-tuned the SFT variant of Zephyr using DPO with our binarized version of UltraFeedback for intent alignment, which resulted in better results for LM Eval Harness (52.89 vs 52.15) and AlpacaEval (91.42 vs 90.60), while almost on par in MT-Bench (7.30 vs 7.34) 🎉
1
1
6
💥 I've just included formatting support for
@GoogleDeepMind
Gemma models within `llama-cpp-python` from
@abetlen
It's as easy as providing the `chat_format=gemma` to the Python client using one of the GGUF quantized versions of the models
See the PR at
0
0
6
Additionally, if you are interested in building your own preference datasets with AI Feedback (AIF), feel free to give a try!
0
1
6
☹️ Time to say bye to the Notus Chat UI hosted in
@huggingface
Spaces!
It's been great to have so many people liking and using it, we got to be the space of the week, and much more, but we're now looking forward what's to come!
1
1
6
My year on
#Github2021
🐙 alvarobartt
📬 Commits/Issues/PRs: 889
🏝️ Repos contributed to: 12
⭐ New stars: 853
🔥 Hottest: alvarobartt/investpy (+507)
Share yours: | Built by
@jrieke
w/
@streamlit
0
1
5
🧐 Did you know that `investiny` is faster, lighter and easier to use than `investpy`?
Check it out at
1
1
5
Zephyr was released by the
@huggingface
H4 team and it topped other 7B models (and bigger) on chat benchmarks, and they also demonstrated that DPO fine-tuning with preference data helps on intent alignment.
1
0
5
tired: phi-2 in azure model catalog
wired: mixtral in a torrent
0
0
4
These were the analytics in case anyone was wondering, great ones! Thanks everyone 🤗
0
1
5
Thanks for the mention to our work at Argilla! And congrats on the nice release 🫶🏻
👑 AlphaMonarch-7B
I'm releasing a family of 7B models that are top performers on many benchmarks, including Nous, EQ-bench, MT-Bench, and the Open LLM Leaderboard.
They also include a few new ideas that are quite promising for the future of 7B models. 🧵
🤗 Model:
19
79
432
0
1
5
Also wanted to point out my very first ML related repository made open source, which is a detailed explanation on how to serve
@PyTorch
ML models in production using TorchServe 🔥
Notice the growth impact after being featured in
@mrdbourke
's video:
1
0
5
🤗 The weights are available in the
@huggingface
Hub, and you can run the inference in any MPS chip with at least 32 GB of RAM
1
0
5
We are so close before we release the initial version of the dataset publicly in the HuggingFace Hub!
Help us reach the milestone so that we can all benefit of the collaborative effort 🤝🏻
Help us build a 10k dataset in just two weeks! Data is Better Together is 2500 annotations away from the goal.
If you have a
@huggingface
account, you can contribute preference labels in under 1 minute.
Join here:
2
2
15
0
2
5
@winglian
@argilla_io
@Teknium1
wow impressive! glad you found our datasets useful and it’s nice to see those results, those only keep on demonstrating how relevant the data quality is! 🤗
0
0
5
@Teknium1
@huggingface
@ByteNano
wow this is great! on resource par with SFT LoRA and QLoRA fine-tuning 👏🏻
0
0
5
@comma_ai
👏🏻 that's how you partner with open-source projects! keep that economic profit bullshit aside and collaborate on projects you'd like to
0
0
4
le café notus anyone?
If you want to predict the next big LLM, just take a stroll around Paris :)
12
16
143
0
1
5
@morgymcg
and
@capetorch
have some compute and are running early 2024 experiments with Mixtral 📢
feel free to propose some sweeps to them to get more insights about Mixtral fine-tuning 👇🏻
Anyone want to suggest some Mixtral 8x7b fine-tuning configs to try?
Myself and
@capetorch
have 8 x H100s (thanks to
@CoreWeave
🙇) for 5 more days and are doing some explorations find a decent Mixtral fine-tuning recipe using axolotl that we can share.
What configs should we
7
20
117
0
2
5
FYI we also created a 🤗 Hub Collection with all the preference datasets that we will be producing in the upcoming months!
1
0
4
@weights_biases
always great to see the loss go down in wandb 🎉 also kudos to
@capetorch
on his Zephyr reproduction! exchanging messages in github and looking at his runs was also really insightful 🫶🏻
p.s. oops made a typo in the mention
@weights_biases
:(
1
0
4
I'm curious about why so many people are following me on GitHub lately? Did anyone promote my profile somewhere? Did I skip any relevant mention?
3
0
3
We’d love to hear some feedback from everyone trying it out, and also kudos to my team mates
@dvilasuero
and
@gabrielmbmb_
as we’ve working really hard to ship this! 🫶🏻
Still on an early stage, but stay tuned because we’re actively working on improving and extending it 🔥
0
0
4
@jrieke
@I_m_Jojo
@GuglielmoIozzia
@gbhatia30
@JudeLeonard13
@JavierFnts
@Rasa_HQ
Thanks for the mention Johannes! 🎈
You can check the
@streamlit
application I created, with TensorFlow Serving as the backend serving a CNN trained to identify some of the main "The Simpsons" characters!!
💻Source code:
1
0
4
@HaihaoShen
@huggingface
@IntelAI
@lvkaokao
this just keeps on showcasing the relevance of the DPO fine-tune right after the SFT one, great job
@HaihaoShen
👏🏻
0
0
4
going to japan in 1 week for about 2 weeks, name something I cannot miss
3
0
4
Additionally, we also aligned a bit the formatting so that anyone willing to use it with
@huggingface
H4 alignment handbook is able to do so with minimal to no extra configuration 🤗
cc
@_lewtun
@edwardbeeching
1
0
4
@awnihannun
@pcuenq
this is great! looking forward trying those out and also uploading some 🫶🏻
1
0
4
Good news for all the investpy users out there! I'm proud to announce that 2 friends will be now investpy collaborators for the upcoming releases of the package and also for newer features! They are
@Juanju97Gates
and
@gmartin_b
🥳
2
0
4
time to fine-tune 👀
0
1
4
🚨 Shutting it down in 1 day! So make sure to try it out before it's not on anymore
🤗 Space at
☹️ Time to say bye to the Notus Chat UI hosted in
@huggingface
Spaces!
It's been great to have so many people liking and using it, we got to be the space of the week, and much more, but we're now looking forward what's to come!
1
1
6
0
2
4
hi
@AnthropicAI
, do you have an estimate date on when will Claude be available in Spain? I'm really looking forward trying it out 🤗
1
0
3
@HamelHusain
thanks for the review
@HamelHusain
! we're improving, and looking forward to see you back using
@argilla_io
, all the best!
0
0
4
cool
@huggingface
hub release today 🎉 the follow button has arrived!
probably
@TheBlokeAI
is a more interesting follow, but you can also follow me at 🤗
0
0
4
Hoy tengo buenas noticias para los usuarios de investpy!! Unos amigos y ahora colaboradores se unirán conmigo en el desarrollo de las versiones sucesivas tanto del paquete como de funcionalidades adicionales. Ellos son
@Juanju97Gates
y
@gmartin_b
1
1
4
👏🏻 gran feina la que està fent la gent del
@projecte_aina
per la inclusió del català dins la intersecció entre NLP i el codi obert
us recomano investigar sobre ells i explorar els recursos oberts en català que tenen disponibles al Hub de
@huggingface
0
0
4
To anyone interested in NLP (NER, SpanMarker, SentenceTransformers, Attention Sinks, or anything really) make sure to follow Tom 🐐
Thrilled to share that I've joined
@huggingface
🤗 as a Machine Learning Engineer tasked with maintaining the awesome Sentence Transformers project!
It's high time to bring modern training functionality to finetuning embedding models!⚡️
5
11
125
0
1
4