How to run inference after SFT tuning ?

[alphaversedev]

Hi,
Is there a sample code of inference after finetuning using SFT ? I can see files are created in the output dir, but got error saying : RuntimeError: You can't move a model that has some modules offloaded to cpu or disk.

this is how I load the model:
model = AutoModelForCausalLM.from_pretrained(
"D:/model_folder",
device_map="auto",
trust_remote_code=True,
temperature = 0.8).eval()

I am using Qwen1.5 for the finetune.

thanks
lixin

[hiyouga]

use CLI demo to infer trained models

[alphaversedev]

thanks, it works, though sometime it keeps generating duplicated answers.
By adding torch_dtype=torch.float16, I am also able to generate answers from the finetuned result now.

I am using Qwen1.5-1.8B-Chat-GPTQ-Int8, with a very small dataset (80+ pairs). the answer does not seem to be good enough. I am not sure if this is the quality/quantity of the dataset, or something else.

[alphaversedev]

at end of the finetuning, saw this :
[INFO|modelcard.py:452] 2024-02-08 17:31:51,344 >> Dropping the following result as it does not have all the necessary fields:
{'task': {'name': 'Causal Language Modeling', 'type': 'text-generation'}}

anything wrong or can just ignore this ?

[hiyouga]

ignore it

[alphaversedev]

I tried increasing the n_epochs, from 3 to eventually 800, finally getting some meaningful response. However, I worry that this will have model overfitting.

The loss is pretty low now :
epoch = 800.0
train_loss = 0.0605

appreciate your work and support very much, @hiyouga !