I tried Llama for it, it works, but it appears to be too expensive. Also tried Gemma, but it didn’t understand the task at all.
I’m curious, is there any known LLMs or even regular models that can do it?
Would appreciate any ideas. Thanks
Both Llama and Gemma come in a variety of sizes (from 0.5B to 72B, and even larger…).
There is a general proportional relationship between size and performance. The size makes a huge difference.
Also, the training status is diverse, and for example, the response to prompts varies considerably between Base and Instruct.
If you want to use it like ChatGPT, I think you should just use Instruct.
So, based on the size of the Llama that worked well and whether it was Instruct or Base, it would be good to try to find a smaller similar LLM.
I can recommend Qwen 2.5 as a relatively high-performance Instruct model, even if it is small. Llama is also excellent, even if it is small, especially after 3.2B. I think Gemma 9B is also excellent, and 2B is also very high-performance for its size. I have not used Base (or models without a specific description may be Base) very actively, so I cannot compare it.
Thanks for sharing your experience! I’ve heard similar feedback about Llama – it’s effective but can be pricey. As for Gemma, yeah, sometimes those models can miss the mark depending on the task. In terms of other models, you could try looking into OpenAI’s GPT models, especially with fine-tuning for specific tasks. Also, models like Cohere and Anthropic’s Claude might be worth checking out; they offer competitive pricing and have been solid for certain use cases. Another option could be Hugging Face’s open models, where you can experiment with different architectures based on your needs. Let me know if any of these sound like they could work for you