File size: 2,638 Bytes
c343e79
 
 
 
 
 
 
 
 
 
ee8408d
620ea14
 
 
c343e79
620ea14
c343e79
620ea14
 
 
4c24792
70fc867
 
620ea14
69fb7e4
 
620ea14
70fc867
620ea14
69fb7e4
620ea14
 
c343e79
69fb7e4
469d285
69fb7e4
469d285
 
 
 
 
69fb7e4
 
 
469d285
 
 
 
 
 
 
69fb7e4
 
c343e79
 
 
 
 
 
620ea14
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
---
language:
- en
license: apache-2.0
tags:
- text-generation-inference
- transformers
- unsloth
- mistral
- trl
- sft
base_model: unsloth/mistral-7b-v0.2-bnb-4bit
datasets:
- visheratin/realworldqa
---
# Mistral-RealworldQA-v0.2-7b SFT

<img src="https://i.imgur.com/Pf53ms5.jpeg" width="400"/>


An experiment with the goal of reducing hallucinations in [VQA](https://huggingface.co/tasks/visual-question-answering)

First in a series of experiments centering around fine-tuning for image captioning.

<h1>Release Notes</h1>

* v0.1 - Initial Release
* <b>v0.2</b> (Current)- Updating base model to official Mistral-7b fp16 release, refinements to dataset and instruction formating

<h2>Background & Methodology</h2>

Mistral-7b-02 base model was fine-tuned using the [RealWorldQA dataset](https://huggingface.co/datasets/visheratin/realworldqa), originally provided by the X.Ai Team here: https://x.ai/blog/grok-1.5v

<h1>Vision Results</h1>
<img src="https://i.imgur.com/E9mS4Xb.jpeg" width="400"/>

* Experiment yielded model that provides shorter, less verbose output for questions about pictures
* The likelihood of hallucinations in output has decreased, however, the model can still be easily influenced to be inaccurate by the user
* Best suited for captioning use cases that require concise descriptions and low token counts
* This model lacks the conversational prose of Excalibur-7b-DPO and is much "drier" in tone

<b>Requires additional mmproj file. You have two options for vision functionality (available inside this repo):</b>
 1. [Quantized - Limited VRAM Option (197mb)](https://huggingface.co/InferenceIllusionist/Excalibur-7b-DPO-GGUF/resolve/main/mistral-7b-mmproj-v1.5-Q4_1.gguf?download=true)
 2. [Unquantized - Premium Option / Best Quality (596mb)](https://huggingface.co/InferenceIllusionist/Excalibur-7b-DPO-GGUF/resolve/main/mmproj-model-f16.gguf?download=true)

Select the gguf file of your choice in [Koboldcpp](https://github.com/LostRuins/koboldcpp/releases/) as usual, then make sure to choose the mmproj file above in the LLaVA mmproj field of the model submenu:
<img src="https://i.imgur.com/x8vqH29.png" width="425"/>

## Prompt Format
Use Alpaca for best results.


## Other info
- **Developed by:** InferenceIllusionist
- **License:** apache-2.0
- **Finetuned from model :** mistral-community/Mistral-7B-v0.2

This mistral model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)