openbmb

/

MiniCPM5-2B

Like 1.49k

Follow

OpenBMB 5.18k

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.

Even though Spark-X2.5's turkish support is better, it is still far from perfect as it sometimes injects chinese characters in the response and also it makes grammatical mistakes with suffixes in turkish.

i want to fine tune this model but i dont have enough resources,i just hope after updates makes this model supports turkish.

I am looking for ways to do it efficiently, I will let you know if I find a way I currently have 24GB of VRAM so it might be feasible on my system but there could be a massive degradation if it is done improperly.

I sincerely hope you can find similar approaches. We truly wish for barrier-free access to cutting-edge technology in more regions.

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.

Even though Spark-X2.5's turkish support is better, it is still far from perfect as it sometimes injects chinese characters in the response and also it makes grammatical mistakes with suffixes in turkish.

i want to fine tune this model but i dont have enough resources,i just hope after updates makes this model supports turkish.

I am looking for ways to do it efficiently, I will let you know if I find a way I currently have 24GB of VRAM so it might be feasible on my system but there could be a massive degradation if it is done improperly.

gemma and qwen models are surely better in turkish (especially gemma

use gemma instead

use gemma instead

i mean gemma is good at turkish but cpm 2b gas better benchmarks,it will be good if there is a model do them at same time

use gemma instead

i mean gemma is good at turkish but cpm 2b gas better benchmarks,it will be good if there is a model do them at same time

cpm 2b is a 2B model it isnt gonna be great with anything. just use the one that can speak your language.

use gemma instead

i mean gemma is good at turkish but cpm 2b gas better benchmarks,it will be good if there is a model do them at same time

cpm 2b is a 2B model it isnt gonna be great with anything. just use the one that can speak your language.

ok,thanks for suggestion

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.

Even though Spark-X2.5's turkish support is better, it is still far from perfect as it sometimes injects chinese characters in the response and also it makes grammatical mistakes with suffixes in turkish.

i want to fine tune this model but i dont have enough resources,i just hope after updates makes this model supports turkish.

I am looking for ways to do it efficiently, I will let you know if I find a way I currently have 24GB of VRAM so it might be feasible on my system but there could be a massive degradation if it is done improperly.

gemma and qwen models are surely better in turkish (especially gemma

Thanks a lot for the feedback! We’ll put more focus on multilingual support~

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.
Spark-X2.5 isn’t a better option for Turkish either.

Dear OpenBMB team, I have a plan to add turkish support to this model via fine-tuning and I want to hear your thoughts about it.
My current plan is to have a %80 turkish %20 english dataset which is going to be a subset of the data used for training this model. (Turkish data is going to be the translation of openbmb/UltraData-SFT-2605, English data is going to be a subset of openbmb/UltraData-RL-2609)
Since the RL+OPD weights are brittle my plan is to start from the SFT checkpoint then whenever Turkish data is picked, I will use plain CE loss and whenever English data is picked I will use the RL+OPD as the teacher and do OPD via Reverse KL on top-32 probs. I am aware this wouldn't be as capable as your RL+OPD weights but this seemed the most viable method. Based off your experience, does this method make sense?

Belki Spark-X2.5'ı deneyebilirsin. Bence çok dilli destek sunuyorlar.

Even though Spark-X2.5's turkish support is better, it is still far from perfect as it sometimes injects chinese characters in the response and also it makes grammatical mistakes with suffixes in turkish.

i want to fine tune this model but i dont have enough resources,i just hope after updates makes this model supports turkish.

I am looking for ways to do it efficiently, I will let you know if I find a way I currently have 24GB of VRAM so it might be feasible on my system but there could be a massive degradation if it is done improperly.

gemma and qwen models are surely better in turkish (especially gemma

Thanks a lot for the feedback! We’ll put more focus on multilingual support~

thanks,it will be sooooo good❤️