SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
Paper
•
2502.02737
•
Published
•
54
SomosNLP
non-profit
AI & ML interests
Democratizar el PLN en español e incentivar su aplicación para generar impacto social 💛
Recent Activity
View all activity
somosnlp's activity
lewtun
authored
a
paper
35 minutes ago
mariagrandury
updated
a
collection
2 days ago
mariagrandury
updated
2
collections
8 days ago
Post
9875
We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open!
🧪 Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.
🧠 Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.
🔥 Step 3: show we can go from base model -> SFT -> RL via multi-stage training.
Follow along: https://github.com/huggingface/open-r1
🧪 Step 1: replicate the R1-Distill models by distilling a high-quality reasoning corpus from DeepSeek-R1.
🧠 Step 2: replicate the pure RL pipeline that DeepSeek used to create R1-Zero. This will involve curating new, large-scale datasets for math, reasoning, and code.
🔥 Step 3: show we can go from base model -> SFT -> RL via multi-stage training.
Follow along: https://github.com/huggingface/open-r1
- 5 replies
Post
3806
I was initially pretty sceptical about Meta's Coconut paper [1] because the largest perf gains were reported on toy linguistic problems. However, these results on machine translation are pretty impressive!
https://x.com/casper_hansen_/status/1875872309996855343
Together with the recent PRIME method [2] for scaling RL, reasoning for open models is looking pretty exciting for 2025!
[1] Training Large Language Models to Reason in a Continuous Latent Space (2412.06769)
[2] https://huggingface.co/blog/ganqu/prime
https://x.com/casper_hansen_/status/1875872309996855343
Together with the recent PRIME method [2] for scaling RL, reasoning for open models is looking pretty exciting for 2025!
[1] Training Large Language Models to Reason in a Continuous Latent Space (2412.06769)
[2] https://huggingface.co/blog/ganqu/prime
Post
2266
This paper (
HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs (2412.18925)) has a really interesting recipe for inducing o1-like behaviour in Llama models:
* Iteratively sample CoTs from the model, using a mix of different search strategies. This gives you something like Stream of Search via prompting.
* Verify correctness of each CoT using GPT-4o (needed because exact match doesn't work well in medicine where there are lots of aliases)
* Use GPT-4o to reformat the concatenated CoTs into a single stream that includes smooth transitions like "hmm, wait" etc that one sees in o1
* Use the resulting data for SFT & RL
* Use sparse rewards from GPT-4o to guide RL training. They find RL gives an average ~3 point boost across medical benchmarks and SFT on this data already gives a strong improvement.
Applying this strategy to other domains could be quite promising, provided the training data can be formulated with verifiable problems!
* Iteratively sample CoTs from the model, using a mix of different search strategies. This gives you something like Stream of Search via prompting.
* Verify correctness of each CoT using GPT-4o (needed because exact match doesn't work well in medicine where there are lots of aliases)
* Use GPT-4o to reformat the concatenated CoTs into a single stream that includes smooth transitions like "hmm, wait" etc that one sees in o1
* Use the resulting data for SFT & RL
* Use sparse rewards from GPT-4o to guide RL training. They find RL gives an average ~3 point boost across medical benchmarks and SFT on this data already gives a strong improvement.
Applying this strategy to other domains could be quite promising, provided the training data can be formulated with verifiable problems!
- 1 reply
Post
6827
We outperform Llama 70B with Llama 3B on hard math by scaling test-time compute 🔥
How? By combining step-wise reward models with tree search algorithms :)
We show that smol models can match or exceed the performance of their much larger siblings when given enough "time to think"
We're open sourcing the full recipe and sharing a detailed blog post.
In our blog post we cover:
📈 Compute-optimal scaling: How we implemented DeepMind's recipe to boost the mathematical capabilities of open models at test-time.
🎄 Diverse Verifier Tree Search (DVTS): An unpublished extension we developed to the verifier-guided tree search technique. This simple yet effective method improves diversity and delivers better performance, particularly at large test-time compute budgets.
🧭 Search and Learn: A lightweight toolkit for implementing search strategies with LLMs and built for speed with vLLM
Here's the links:
- Blog post: HuggingFaceH4/blogpost-scaling-test-time-compute
- Code: https://github.com/huggingface/search-and-learn
Enjoy!
How? By combining step-wise reward models with tree search algorithms :)
We show that smol models can match or exceed the performance of their much larger siblings when given enough "time to think"
We're open sourcing the full recipe and sharing a detailed blog post.
In our blog post we cover:
📈 Compute-optimal scaling: How we implemented DeepMind's recipe to boost the mathematical capabilities of open models at test-time.
🎄 Diverse Verifier Tree Search (DVTS): An unpublished extension we developed to the verifier-guided tree search technique. This simple yet effective method improves diversity and delivers better performance, particularly at large test-time compute budgets.
🧭 Search and Learn: A lightweight toolkit for implementing search strategies with LLMs and built for speed with vLLM
Here's the links:
- Blog post: HuggingFaceH4/blogpost-scaling-test-time-compute
- Code: https://github.com/huggingface/search-and-learn
Enjoy!
- 2 replies
Post
567
🌐 The Stanford Institute for Human-Centered AI (https://aiindex.stanford.edu/vibrancy/) has released its 2024 Global AI Vibrancy Tool, a way to explore and compare AI progress across 36 countries.
📊 It measures progress across the 8 broad pillars of R&D, Responsible AI, Economy, Education, Diversity, Policy and Governance, Public Opinion and Infrastructure. (Each of these pillars have a number of Sub Indices)
📈 As a whole it is not surprising that the USA was at the top in terms of overall score as of 2023 (AI investment activity is a large part of the economic pillar for example and that is a large part of the overall USA ranking) but drilling in to more STRATEGIC Macro pillars like Education, Infrastructure or R&D reveal interesting growth patterns in Asia (particularly China) and Western Europe that I suspect the 2024 metrics will bear out.
🤖 Hopefully the 2024 Global Vibrancy ranking will break out AI and ML verticals like Computer Vision or NLP and or the AI Agent space as that may also from a global macro level give indications of what is to come globally for AI in 2025.
📊 It measures progress across the 8 broad pillars of R&D, Responsible AI, Economy, Education, Diversity, Policy and Governance, Public Opinion and Infrastructure. (Each of these pillars have a number of Sub Indices)
📈 As a whole it is not surprising that the USA was at the top in terms of overall score as of 2023 (AI investment activity is a large part of the economic pillar for example and that is a large part of the overall USA ranking) but drilling in to more STRATEGIC Macro pillars like Education, Infrastructure or R&D reveal interesting growth patterns in Asia (particularly China) and Western Europe that I suspect the 2024 metrics will bear out.
🤖 Hopefully the 2024 Global Vibrancy ranking will break out AI and ML verticals like Computer Vision or NLP and or the AI Agent space as that may also from a global macro level give indications of what is to come globally for AI in 2025.
Post
819
🤖💻 Function Calling is a key component of Agent workflows. To call functions, an LLM needs a way to interact with other systems and run code. This usually means connecting it to a runtime environment that can handle function calls, data, and security.
Per the Berkeley Function-Calling Leaderboard there are only 2 fully open source models (The other 2 in the top 20 that are not closed source have cc-by-nc-4.0 licenses) out of the top 20 models that currently have function calling built in as of 17 Nov 2024.
https://gorilla.cs.berkeley.edu/leaderboard.html
The 2 Open Source Models out of the top 20 that currently support function calling are:
meetkai/functionary-medium-v3.1
Team-ACE/ToolACE-8B
This is a both a huge disadvantage AND an opportunity for the Open Source community as Enterprises, Small Business, Government Agencies etc. quickly adopt Agents and Agent workflows over the next few months. Open Source will have a lot of catching up to do as Enterprises will be hesitant to switch from the closed source models that they may initially build their Agent workflows on in the next few months to an open source alternative later.
Hopefully more open source models will support function calling in the near future.
Per the Berkeley Function-Calling Leaderboard there are only 2 fully open source models (The other 2 in the top 20 that are not closed source have cc-by-nc-4.0 licenses) out of the top 20 models that currently have function calling built in as of 17 Nov 2024.
https://gorilla.cs.berkeley.edu/leaderboard.html
The 2 Open Source Models out of the top 20 that currently support function calling are:
meetkai/functionary-medium-v3.1
Team-ACE/ToolACE-8B
This is a both a huge disadvantage AND an opportunity for the Open Source community as Enterprises, Small Business, Government Agencies etc. quickly adopt Agents and Agent workflows over the next few months. Open Source will have a lot of catching up to do as Enterprises will be hesitant to switch from the closed source models that they may initially build their Agent workflows on in the next few months to an open source alternative later.
Hopefully more open source models will support function calling in the near future.
mariagrandury
updated
a
collection
3 months ago
Post
2278
The Mystery Bot 🕵️♂️ saga I posted about from earlier this week has been solved...🤗
Cohere for AI has just announced its open source Aya Expanse multilingual model. The Initial release supports 23 languages with more on the way soon.🌌 🌍
You can also try Aya Expanse via SMS on your mobile phone using the global WhatsApp number or one of the initial set of country specific numbers listed below.⬇️
🌍WhatsApp - +14313028498
Germany - (+49) 1771786365
USA – +18332746219
United Kingdom — (+44) 7418373332
Canada – (+1) 2044107115
Netherlands – (+31) 97006520757
Brazil — (+55) 11950110169
Portugal – (+351) 923249773
Italy – (+39) 3399950813
Poland - (+48) 459050281
Cohere for AI has just announced its open source Aya Expanse multilingual model. The Initial release supports 23 languages with more on the way soon.🌌 🌍
You can also try Aya Expanse via SMS on your mobile phone using the global WhatsApp number or one of the initial set of country specific numbers listed below.⬇️
🌍WhatsApp - +14313028498
Germany - (+49) 1771786365
USA – +18332746219
United Kingdom — (+44) 7418373332
Canada – (+1) 2044107115
Netherlands – (+31) 97006520757
Brazil — (+55) 11950110169
Portugal – (+351) 923249773
Italy – (+39) 3399950813
Poland - (+48) 459050281
- 1 reply
Post
2521
Spent the weekend testing out some prompts with 🕵️♂️Mystery Bot🕵️♂️ on my mobile... exciting things are coming soon for the following languages:
🌐Arabic, Chinese, Czech, Dutch, English French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese!🌐
🌐Arabic, Chinese, Czech, Dutch, English French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese!🌐
mariagrandury
authored
a
paper
4 months ago
Post
1392
📢 2024 CVPR Videos Are Now Available! 🎥
CVPR conference keynotes, panels, posters, workshops, and other content are now available.
⬇️
https://cvpr.thecvf.com/Conferences/2024/Videos
CVPR conference keynotes, panels, posters, workshops, and other content are now available.
⬇️
https://cvpr.thecvf.com/Conferences/2024/Videos
Post
2352
💡Andrew Ng recently gave a strong defense of Open Source AI models and the need to slow down legislative efforts in the US and the EU to restrict innovation in Open Source AI at Stanford GSB.
🎥See video below
https://youtu.be/yzUdmwlh1sQ?si=bZc690p8iubolXm_
🎥See video below
https://youtu.be/yzUdmwlh1sQ?si=bZc690p8iubolXm_
- 4 replies
daqc
updated
a
model
6 months ago
mariagrandury
updated
a
Space
6 months ago
mariagrandury
authored
a
paper
7 months ago
Post
810
Researchers from Auburn University and the University of Alberta have explored the limitations of Vision Language Models (VLMs) in their recently published paper titled "Vision language models are blind." (
Vision language models are blind (2407.06581))
Key Findings:🔍
VLMs, including GPT-4o, Gemini-1.5 Pro, Claude-3 Sonnet, and Claude-3.5 Sonnet, struggle with basic visual tasks.
Tasks such as identifying where lines intersect or counting basic shapes are challenging for these models.
The authors noted, "The shockingly poor performance of four state-of-the-art VLMs suggests their vision is, at best, like of a person with myopia seeing fine details as blurry, and at worst, like an intelligent person that is blind making educated guesses"(Vision Language Models Are Blind; 2024).
Human-like Myopia? 👓
VLMs may have a blind spot similar to human myopia.
This limitation makes it difficult for VLMs to perceive details.
Suggests a potential parallel between human and machine vision limitations.
Technical Details: 🔧
The researchers created a new benchmark called BlindTest.
BlindTest consists of simple visual tasks to evaluate VLMs low-level vision capabilities.
Four VLMs were assessed using BlindTest.
Many shortcomings were revealed in the models ability to process basic visual information.
Learn More: 🖼️
For a deeper dive into this research, check out the project page: https://vlmsareblind.github.io/
Key Findings:🔍
VLMs, including GPT-4o, Gemini-1.5 Pro, Claude-3 Sonnet, and Claude-3.5 Sonnet, struggle with basic visual tasks.
Tasks such as identifying where lines intersect or counting basic shapes are challenging for these models.
The authors noted, "The shockingly poor performance of four state-of-the-art VLMs suggests their vision is, at best, like of a person with myopia seeing fine details as blurry, and at worst, like an intelligent person that is blind making educated guesses"(Vision Language Models Are Blind; 2024).
Human-like Myopia? 👓
VLMs may have a blind spot similar to human myopia.
This limitation makes it difficult for VLMs to perceive details.
Suggests a potential parallel between human and machine vision limitations.
Technical Details: 🔧
The researchers created a new benchmark called BlindTest.
BlindTest consists of simple visual tasks to evaluate VLMs low-level vision capabilities.
Four VLMs were assessed using BlindTest.
Many shortcomings were revealed in the models ability to process basic visual information.
Learn More: 🖼️
For a deeper dive into this research, check out the project page: https://vlmsareblind.github.io/