Generation of response not stopping, keeps "talking" with itself.

#8
by daniq - opened

Hello.

I've been trying to use this model for writing summaries. I've loaded it into oobabooga text generation web ui, then on default page made this prompt:

You are an AI assistant that converts complicated image descriptions into compiled descriptive captions.

### Instruction:
1. Caption should be in third person.
2. Caption should be single sentence.
3. Do not use word "Character".

### Input:
Description: The image showcases three animated characters in a room with bookshelves in the background. The character on the left has black hair and is smiling, the one in the center has red hair and is looking surprised, and the character on the right has purple hair and is partially visible.

### Response:

When I click on generate, it goes insane. It start generate texts without stopping and keeps "talking" with itself until it exhausts all it's max_new_tokens param value.

Sometimes it breaks completely and starts writing giant texts on Korean.

I've loaded this model into oobabooga with default settings.

NousResearch org

Hello.

I've been trying to use this model for writing summaries. I've loaded it into oobabooga text generation web ui, then on default page made this prompt:

You are an AI assistant that converts complicated image descriptions into compiled descriptive captions.

### Instruction:
1. Caption should be in third person. 
2. Caption should be single sentence.
3. Do not use word "Character".

### Input:
Description: The image showcases three animated characters in a room with bookshelves in the background. The character on the left has black hair and is smiling, the one in the center has red hair and is looking surprised, and the character on the right has purple hair and is partially visible.

### Response:

When I click on generate, it goes insane. It start generate texts without stopping and keeps "talking" with itself until it exhausts all it's max_new_tokens param value.

Sometimes it breaks completely and starts writing giant texts on Korean.

I've loaded this model into oobabooga with default settings.

You're using the wrong prompt format. See the model card to understand the format, it is chatml, not alpaca.

teknium changed discussion status to closed

@teknium I thought that default page of text generation web ui already includes this prompt format implicitly. Thank you.

Sign up or log in to comment