haotian-liu / LLaVA

[NeurIPS'23 Oral] Visual Instruction Tuning (LLaVA) built towards GPT-4V level capabilities and beyond.
https://llava.hliu.cc
Apache License 2.0
20.46k stars 2.26k forks source link

[Question] The results of the local model are inconsistent with the web ui in the demo #1497

Open zmf2022 opened 6 months ago

zmf2022 commented 6 months ago

Question

May I ask, based on the same test data, using the llava: 34b-v1.6-q3_K_S model locally and the llava v1.6-34b model for web UI in the demo, there is a significant difference in the results obtained between the two. The results in web UI are significantly better than those in the local environment. What is the reason for this? The problem is: Which country recorded the highest death rates due to outdoor air pollution over the years? The result in the web UI is: The graph you've provided shows the outdoor air pollution death rate for several countries from 2005 to 2012. The country with the highest death rate due to outdoor air pollution over the years shown on the graph is Myanmar. The line representing Myanmar's death rate is the highest among the countries depicted, indicating a higher number of deaths attributed to outdoor air pollution per 100,000 people. The local result is:India recorded the highest death rates due to outdoor air pollution over the years according to this chart ![Uploading 00795994017065.png…]()

avalonliberty commented 6 months ago

I suffered the issue here. The web demo version ran this "llava-v1.6-34b" version of llava. Given the "34b-v1.6" version of model in Ollama, the results of queries into these two models are significantly different as well. The one in the web demo outperformed the one in ollama. May I know which tag in Ollama can match the performance in the web demo?

ChristianWeyer commented 6 months ago

I am seeing the exact same thing.

Currently, I am running Llava:34b-V1.6 in Ollama, and it just does not recognize things in a form and hallucinates values - whereas the HF-hosted Llava 34b (https://llava.hliu.cc) does a great job.

Who should we ping here? .cc @haotian-liu Thanks!