nintwentydo's picture
Update README.md
3683ce6 verified
metadata
license: llama3.2
base_model:
  - meta-llama/Llama-3.2-11B-Vision-Instruct
  - invisietch/L3.1-EtherealRainbow-v1.0-rc1-8B
base_model_relation: merge
language:
  - en
pipeline_tag: image-text-to-text
tags:
  - not-for-all-audiences
  - llama
  - llama-3
  - vision
  - multimodal

Quick experiment merging Llama 3.2 11B Vision adapters onto a Llama 3.1 finetune using Grimulkan's fantastic script outlined here.

Vision adapters: meta-llama/Llama-3.2-11B-Vision
Language model: invisietch/L3.1-EtherealRainbow-v1.0-rc1-8B

If you have issues with the tokenizer use the Llama 3.2 11B tokenizer.

If you want to use system messages with image inputs use the template in chat_template.json. No guarantees it'll work properly but seems to be ok for me so far.

Sampling params seem to be a bit finnicky, currently testing out different settings and would love to hear what works for y'all.