QWEN CHAT Hugging Face ModelScope DEMO DISCORD
Scaling Reinforcement Learning (RL) has the potential to enhance model performance beyond conventional pretraining and post-training methods. Recent studies have demonstrated that RL can significantly improve the reasoning capabilities of models. For instance, DeepSeek R1 has achieved state-of-the-art performance by integrating cold-start data and multi-stage training, enabling deep thinking and complex reasoning.
Our research explores the scalability of Reinforcement Learning (RL) and its impact on enhancing the intelligence of large language models.
Ultimately what matters is whether it gets the correct answer or not. It’s interesting that yours wasn’t able to do the strawberry test while mine did it with very short thinking cycle.
Ultimately what matters is whether it gets the correct answer or not.
That’s… not true at all. It had the right answer, to most of the questions I asked it, just as fast as R1, and yet it kept saying “but wait! maybe I’m wrong”. It’s a huge red flag when the CoT is just trying to 1000 monkeys a problem.
While it did manage to complete the strawberry problem when I adjusted the top_p/top_k, I was using the previous values with other models I’ve tested and never had a CoT go that off kilter before. And this is considering even the 7B Deepseek model was able to get the correct answer for 1/4 of the vram.
It’s true for me. I generally don’t read through the think part. I make the query, do something else, and then come back to see what the actual output it. Overall, I find it gives me way better answers than I got with the version of R1 I was able to get running locally. Turns out the settings do matter though.
Ultimately what matters is whether it gets the correct answer or not. It’s interesting that yours wasn’t able to do the strawberry test while mine did it with very short thinking cycle.
That’s… not true at all. It had the right answer, to most of the questions I asked it, just as fast as R1, and yet it kept saying “but wait! maybe I’m wrong”. It’s a huge red flag when the CoT is just trying to 1000 monkeys a problem.
While it did manage to complete the strawberry problem when I adjusted the top_p/top_k, I was using the previous values with other models I’ve tested and never had a CoT go that off kilter before. And this is considering even the 7B Deepseek model was able to get the correct answer for 1/4 of the vram.
It’s true for me. I generally don’t read through the think part. I make the query, do something else, and then come back to see what the actual output it. Overall, I find it gives me way better answers than I got with the version of R1 I was able to get running locally. Turns out the settings do matter though.