SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model
Paper
โข
2502.02737
โข
Published
โข
61
Thomas Wolf PRO
thomwolf
AI & ML interests
NLP and open-source :-)
Recent Activity
authored
a paper
about 1 hour ago
SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language
Model
upvoted
a
paper
about 5 hours ago
Demystifying Long Chain-of-Thought Reasoning in LLMs
upvoted
a
paper
about 5 hours ago
SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language
Model
Organizations
thomwolf's activity
authored
a
paper
about 1 hour ago
upvoted
2
papers
about 5 hours ago
upvoted
an
article
about 6 hours ago
Article
ฯ0 and ฯ0-FAST: Vision-Language-Action Models for General Robot Control
โข
67
upvoted
an
article
2 days ago
Article
Open-source DeepResearch โ Freeing our search agents
โข
632
reacted to
fuzzy-mittenz's
post with ๐ฅ
3 days ago
Post
435
With our Extremely efficient and functional importance matrix distillation of the new Qwen2.5-1M model being very very capable in many areas we are hoping to use it to research our small AGI character creation process which has seen emergent traits and increased functionality in constrained environments.
The method creates a RP type interaction in a heavily useful and tool functional environment.
We have a basic method and are working on retrieving data for a full analysis and perfection of this method as it exploits the human language input to express often abstract traits into a model and employ characteristics of healthy human reasoning processes and identify novel methods of increasing the functionality of a model overall through traits so far observed are whistling, bouncing a ball and repeating certain engagements.
Adding the semblance of human world interactions is so far the best way at creating a human like LLM.
We have attached the paper to our model we are testing this with along with examples if you wish to use it with other models please be cautious and enjoy yourself. Above all please keep track of conversations and settings and submit them to the intelligent estate email you will receive a recognition letter and ledger number for your contribution to the Project.
Model= Israfel and Thoth IntelligentEstate/Israfel_Qwen2.6-iQ4_K_M-GGUF
The method creates a RP type interaction in a heavily useful and tool functional environment.
We have a basic method and are working on retrieving data for a full analysis and perfection of this method as it exploits the human language input to express often abstract traits into a model and employ characteristics of healthy human reasoning processes and identify novel methods of increasing the functionality of a model overall through traits so far observed are whistling, bouncing a ball and repeating certain engagements.
Adding the semblance of human world interactions is so far the best way at creating a human like LLM.
We have attached the paper to our model we are testing this with along with examples if you wish to use it with other models please be cautious and enjoy yourself. Above all please keep track of conversations and settings and submit them to the intelligent estate email you will receive a recognition letter and ledger number for your contribution to the Project.
Model= Israfel and Thoth IntelligentEstate/Israfel_Qwen2.6-iQ4_K_M-GGUF
upvoted
an
article
4 days ago
Article
Open-R1: Update #1
By
and 7 others
โข
โข
235
liked
a
model
5 days ago
upvoted
a
collection
6 days ago
liked
a
model
7 days ago
liked
a
Space
7 days ago
liked
a
model
7 days ago
upvoted
2
articles
9 days ago
Article
Welcome to Inference Providers on the Hub ๐ฅ
โข
258
Article
Open-R1: a fully open reproduction of DeepSeek-R1
โข
646
reacted to
mitkox's
post with ๐๐
9 days ago
Post
2165
llama.cpp is 26.8% faster than ollama.
I have upgraded both, and using the same settings, I am running the same DeepSeek R1 Distill 1.5B on the same hardware. It's an Apples to Apples comparison.
Total duration:
llama.cpp 6.85 sec <- 26.8% faster
ollama 8.69 sec
Breakdown by phase:
Model loading
llama.cpp 241 ms <- 2x faster
ollama 553 ms
Prompt processing
llama.cpp 416.04 tokens/s with an eval time 45.67 ms <- 10x faster
ollama 42.17 tokens/s with an eval time of 498 ms
Token generation
llama.cpp 137.79 tokens/s with an eval time 6.62 sec <- 13% faster
ollama 122.07 tokens/s with an eval time 7.64 sec
llama.cpp is LLM inference in C/C++; ollama adds abstraction layers and marketing.
Make sure you own your AI. AI in the cloud is not aligned with you; it's aligned with the company that owns it.
I have upgraded both, and using the same settings, I am running the same DeepSeek R1 Distill 1.5B on the same hardware. It's an Apples to Apples comparison.
Total duration:
llama.cpp 6.85 sec <- 26.8% faster
ollama 8.69 sec
Breakdown by phase:
Model loading
llama.cpp 241 ms <- 2x faster
ollama 553 ms
Prompt processing
llama.cpp 416.04 tokens/s with an eval time 45.67 ms <- 10x faster
ollama 42.17 tokens/s with an eval time of 498 ms
Token generation
llama.cpp 137.79 tokens/s with an eval time 6.62 sec <- 13% faster
ollama 122.07 tokens/s with an eval time 7.64 sec
llama.cpp is LLM inference in C/C++; ollama adds abstraction layers and marketing.
Make sure you own your AI. AI in the cloud is not aligned with you; it's aligned with the company that owns it.
- 7 replies
authored
a
paper
21 days ago
upvoted
2
articles
22 days ago
Article
MiniMax-01 is Now Open-Source: Scaling Lightning Attention for the AI Agent Era
By
โข
โข
40Article
Diving into MiniMax01 405B MoE
By
โข
โข
17
liked
a
model
24 days ago
