Post
1794
everywhere i go i see his face
alkinun
AtAndDev
AI & ML interests
LLMs, Alignment, Merging, Unsloth, DPO, SFT, ORPO, SPIN..
Recent Activity
posted
an
update
5 days ago
everywhere i go i see his face
reacted
to
prithivMLmods's
post
with π
5 days ago
Deepswipe by
.
.
.
. Deepseekπ¬πΏ
Everything is now in recovery. ππ
reacted
to
prithivMLmods's
post
with π₯
5 days ago
Deepswipe by
.
.
.
. Deepseekπ¬πΏ
Everything is now in recovery. ππ
Organizations
AtAndDev's activity
posted
an
update
5 days ago
reacted to
prithivMLmods's
post with ππ₯
5 days ago
Post
4787
Deepswipe by
.
.
.
. Deepseekπ¬πΏ
Everything is now in recovery. ππ
.
.
.
. Deepseekπ¬πΏ
Everything is now in recovery. ππ
- 4 replies
reacted to
onekq's
post with π
9 days ago
Post
2260
So πDeepSeekπ hits the mainstream media. But it has been a star in our little cult for at least 6 months. Its meteoric success is not overnight, but two years in the making.
To learn their history, just look at their π€ repo https://huggingface.co/deepseek-ai
* End of 2023, they launched the first model (pretrained by themselves) following Llama 2 architecture
* June 2024, v2 (MoE architecture) surpassed Gemini 1.5, but behind Mistral
* September, v2.5 surpassed GPT 4o mini
* December, v3 surpassed GPT 4o
* Now R1 surpassed o1
Most importantly, if you think DeepSeek success is singular and unrivaled, that's WRONG. The following models are also near or equal the o1 bar.
* Minimax-01
* Kimi k1.5
* Doubao 1.5 pro
To learn their history, just look at their π€ repo https://huggingface.co/deepseek-ai
* End of 2023, they launched the first model (pretrained by themselves) following Llama 2 architecture
* June 2024, v2 (MoE architecture) surpassed Gemini 1.5, but behind Mistral
* September, v2.5 surpassed GPT 4o mini
* December, v3 surpassed GPT 4o
* Now R1 surpassed o1
Most importantly, if you think DeepSeek success is singular and unrivaled, that's WRONG. The following models are also near or equal the o1 bar.
* Minimax-01
* Kimi k1.5
* Doubao 1.5 pro
- 1 reply
i believe sglang would be even faster but not sure if it supports non-nvidia devices
reacted to
chansung's
post with π₯
11 days ago
Post
1991
Simple Summarization on DeepSeek-R1 from DeepSeek AI
The RL stage is very important.
β³ However, it is difficult to create a truly helpful AI for people solely through RL.
β³ So, we applied a learning pipeline consisting of four stages: providing a good starting point, reasoning RL, SFT, and safety RL, and achieved performance comparable to o1.
β³ Simply fine-tuning other open models with the data generated by R1-Zero (distillation) resulted in performance comparable to o1-mini.
Of course, this is just a brief overview and may not be of much help. All models are accessible on Hugging Face, and the paper can be read through the GitHub repository.
Model: https://huggingface.co/deepseek-ai
Paper: https://github.com/deepseek-ai/DeepSeek-R1
The RL stage is very important.
β³ However, it is difficult to create a truly helpful AI for people solely through RL.
β³ So, we applied a learning pipeline consisting of four stages: providing a good starting point, reasoning RL, SFT, and safety RL, and achieved performance comparable to o1.
β³ Simply fine-tuning other open models with the data generated by R1-Zero (distillation) resulted in performance comparable to o1-mini.
Of course, this is just a brief overview and may not be of much help. All models are accessible on Hugging Face, and the paper can be read through the GitHub repository.
Model: https://huggingface.co/deepseek-ai
Paper: https://github.com/deepseek-ai/DeepSeek-R1
- 1 reply
replied to
nroggendorff's
post
11 days ago
some ppl cant just get enough
reacted to
ezgikorkmaz's
post with ππ
11 days ago
Post
1651
If you are interested in reinforcement learning, a recent paper I wrote introduces foundational analysis on deep reinforcement learning decision making and representations learnt by it.
Link: https://bsky.app/profile/ezgikorkmaz.bsky.social/post/3lfpgsrn6sc2m
Link: https://bsky.app/profile/ezgikorkmaz.bsky.social/post/3lfpgsrn6sc2m
- 1 reply
reacted to
sharpenb's
post with π
11 days ago
Post
1851
We compressed SmolLMs to make 135 variations of them (see https://huggingface.co/PrunaAI?search_models=smolLM) with different quantization configurations with
We made a blog to summarize our findings (see https://www.pruna.ai/blog/smollm2-smaller-faster) and small LM can be made smaller! :)
pruna (https://docs.pruna.ai/en/latest/). We made a blog to summarize our findings (see https://www.pruna.ai/blog/smollm2-smaller-faster) and small LM can be made smaller! :)
-
- 3 replies
That non centered emoji...
But cool blog
reacted to
sometimesanotion's
post with ππ₯
12 days ago
Post
2691
I've managed a #1 score of 41.22% average for 14B parameter models on the Open LLM Leaderboard. As of this writing, sometimesanotion/Lamarck-14B-v0.7 is #8 for all models up to 70B parameters.
It took a custom toolchain around Arcee AI's mergekit to manage the complex merges, gradients, and LoRAs required to make this happen. I really like seeing features of many quality finetunes in one solid generalist model.
It took a custom toolchain around Arcee AI's mergekit to manage the complex merges, gradients, and LoRAs required to make this happen. I really like seeing features of many quality finetunes in one solid generalist model.
- 8 replies
reacted to
onekq's
post with π₯
12 days ago
Post
2647
This is historical. π
DeepSeek πR1π surpassed OpenAI πo1π on the dual leaderboard. What a year for the open source!
onekq-ai/WebApp1K-models-leaderboard
DeepSeek πR1π surpassed OpenAI πo1π on the dual leaderboard. What a year for the open source!
onekq-ai/WebApp1K-models-leaderboard
reacted to
onekq's
post with π₯
12 days ago
Post
4654
πDeepSeek π is the real OpenAI π―
-
- 6 replies
reacted to
chansung's
post with ππ
12 days ago
Post
1991
Simple Summarization on DeepSeek-R1 from DeepSeek AI
The RL stage is very important.
β³ However, it is difficult to create a truly helpful AI for people solely through RL.
β³ So, we applied a learning pipeline consisting of four stages: providing a good starting point, reasoning RL, SFT, and safety RL, and achieved performance comparable to o1.
β³ Simply fine-tuning other open models with the data generated by R1-Zero (distillation) resulted in performance comparable to o1-mini.
Of course, this is just a brief overview and may not be of much help. All models are accessible on Hugging Face, and the paper can be read through the GitHub repository.
Model: https://huggingface.co/deepseek-ai
Paper: https://github.com/deepseek-ai/DeepSeek-R1
The RL stage is very important.
β³ However, it is difficult to create a truly helpful AI for people solely through RL.
β³ So, we applied a learning pipeline consisting of four stages: providing a good starting point, reasoning RL, SFT, and safety RL, and achieved performance comparable to o1.
β³ Simply fine-tuning other open models with the data generated by R1-Zero (distillation) resulted in performance comparable to o1-mini.
Of course, this is just a brief overview and may not be of much help. All models are accessible on Hugging Face, and the paper can be read through the GitHub repository.
Model: https://huggingface.co/deepseek-ai
Paper: https://github.com/deepseek-ai/DeepSeek-R1
-
- 1 reply
reacted to
chansung's
post with π
12 days ago
Post
1981
Simple summarization of Evolving Deeper LLM Thinking (Google DeepMind)
The process starts by posing a question.
1) The LLM generates initial responses.
2) These generated responses are evaluated according to specific criteria (program-based checker).
3) The LLM critiques the evaluated results.
4) The LLM refines the responses based on the evaluation, critique, and original responses.
The refined response is then fed back into step 2). If it meets the criteria, the process ends. Otherwise, the algorithm generates more responses based on the refined ones (with some being discarded, some remaining, and some responses potentially being merged).
Through this process, it demonstrated excellent performance in complex scheduling problems (travel planning, meeting scheduling, etc.). It's a viable method for finding highly effective solutions in specific scenarios.
However, there are two major drawbacks:
π€ An excessive number of API calls are required. (While the cost might not be very high, it leads to significant latency.)
π€ The evaluator is program-based. (This limits its use as a general method. It could potentially be modified/implemented using LLM as Judge, but that would introduce additional API costs for evaluation.)
https://arxiv.org/abs/2501.09891
The process starts by posing a question.
1) The LLM generates initial responses.
2) These generated responses are evaluated according to specific criteria (program-based checker).
3) The LLM critiques the evaluated results.
4) The LLM refines the responses based on the evaluation, critique, and original responses.
The refined response is then fed back into step 2). If it meets the criteria, the process ends. Otherwise, the algorithm generates more responses based on the refined ones (with some being discarded, some remaining, and some responses potentially being merged).
Through this process, it demonstrated excellent performance in complex scheduling problems (travel planning, meeting scheduling, etc.). It's a viable method for finding highly effective solutions in specific scenarios.
However, there are two major drawbacks:
π€ An excessive number of API calls are required. (While the cost might not be very high, it leads to significant latency.)
π€ The evaluator is program-based. (This limits its use as a general method. It could potentially be modified/implemented using LLM as Judge, but that would introduce additional API costs for evaluation.)
https://arxiv.org/abs/2501.09891
reacted to
JingzeShi's
post with π₯
12 days ago
Post
1631
π€©warmup -> stable -> decay leanring rate scheduler:
πuse the Stable Phase CheckPoints to Continue Training the model on Any New Dataset without spikes of the training!!!
SmallDoge/Doge-20M-checkpoint
SmallDoge/Doge-60M-checkpoint
πuse the Stable Phase CheckPoints to Continue Training the model on Any New Dataset without spikes of the training!!!
SmallDoge/Doge-20M-checkpoint
SmallDoge/Doge-60M-checkpoint
-
- 4 replies