search

janhq / jan

Jan is an open source alternative to ChatGPT that runs 100% offline on your computer. Multiple engine support (llama.cpp, TensorRT-LLM)
https://jan.ai/
GNU Affero General Public License v3.0
23.3k stars 1.35k forks source link

bug: LlaVa 7B responds as though I've provided JSON data #3658

Closed erayerdin closed 5 days ago

erayerdin commented 1 month ago

Jan version

0.5.3

Describe the Bug

I've downloaded LlaVa 7B and, to test it, I've sent a simple hello message, and it responded with:

The user provided a JSON object containing a "name" field and a "type" field. The "name" field is a string, while the "type" field is a string as well. The "name" field contains the word "Rock". The "type" field contains the word "type".

Based on the provided JSON object, the "name" field contains a single word and the "type" field also contains a single word. The "name" field is a string, and the "type" field is a string. The JSON object is a valid representation of two pieces of data, a name and a type, but without additional context, it is difficult to determine the specific use case for this data.

Steps to Reproduce

  1. Download LlaVa 7B
  2. Write a message.
  3. Get the weird response.

Screenshots / Logs

image

What is your OS?

imtuyethan commented 1 month ago

Tested on:

Linux:

107 (test-ubuntu-app-cpu-2)

Mac:

Operating System: MacOS Sonoma 14.2 Processor: Apple M2 RAM: 16GB

Results:

On Linux: Screenshot 2024-09-18 at 8 25 25 PM

On Mac: Its responses vary & not stable:

Screenshot 2024-09-18 at 8 32 30 PM Screenshot 2024-09-18 at 8 32 54 PM

High chances are this could be because of model performance, or our model settings are outdated? cc @louis-jan

imtuyethan commented 2 weeks ago

Related: https://github.com/janhq/models/issues/46

imtuyethan commented 5 days ago

I'll close this issue and follow up on https://github.com/janhq/models/issues/46 instead