Commit
•
189b3d3
1
Parent(s):
e2dd0f6
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,61 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: apache-2.0
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
base_model:
|
3 |
+
- InferenceIllusionist/Excalibur-7b
|
4 |
+
library_name: transformers
|
5 |
+
tags:
|
6 |
+
- finetune
|
7 |
+
- dpo
|
8 |
+
- chatml
|
9 |
+
- gguf
|
10 |
+
- imat
|
11 |
license: apache-2.0
|
12 |
+
datasets:
|
13 |
+
- Intel/orca_dpo_pairs
|
14 |
---
|
15 |
+
|
16 |
+
|
17 |
+
# Excalibur-7b-DPO-iMat-GGUF
|
18 |
+
|
19 |
+
<img src="https://i.imgur.com/pbPbqq0.jpeg" width="550"/>
|
20 |
+
|
21 |
+
Quantized from fp32 with love.
|
22 |
+
|
23 |
+
iMatrix .dat file was calculated using groups_merged.txt.
|
24 |
+
|
25 |
+
<b>FP16 available [here](https://huggingface.co/InferenceIllusionist/Excalibur-7b-DPO)</b>
|
26 |
+
|
27 |
+
An initial foray into the world of fine-tuning. The goal of this release was to amplify the quality of the original model's responses, in particular for vision use cases*
|
28 |
+
|
29 |
+
## Notes & Methodology
|
30 |
+
* [Excalibur-7b](https://huggingface.co/InferenceIllusionist/Excalibur-7b) fine-tuned with Direct Preference Optimization (DPO) using Intel/orca_dpo_pairs
|
31 |
+
* This is a quick experiment to determine the impact of DPO finetuning on the original base model
|
32 |
+
* Ran for a little over an hour on a single A100
|
33 |
+
* Internal benchmarks showed improvement over base model, awaiting final results
|
34 |
+
* Precision: bfloat16
|
35 |
+
|
36 |
+
|
37 |
+
## Sample Question - Vision
|
38 |
+
<img src="https://i.imgur.com/7aRWtzU.jpeg" width="425"/>
|
39 |
+
|
40 |
+
*<b>Requires additional mmproj file. You have two options for vision functionality (available inside original repo or linked below):</b>
|
41 |
+
* [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)
|
42 |
+
* [Unquantized - Premium Option / Best Quality (596mb)](https://huggingface.co/InferenceIllusionist/Excalibur-7b-DPO-GGUF/resolve/main/mmproj-model-f16.gguf?download=true)
|
43 |
+
|
44 |
+
Select the gguf file of your choice in Kobold as usual, then make sure to choose the mmproj file above in the LLaVA mmproj field of the model submenu:
|
45 |
+
<img src="https://i.imgur.com/x8vqH29.png" width="425"/>
|
46 |
+
|
47 |
+
## Prompt Format
|
48 |
+
* For best results please use ChatML for the prompt format. Alpaca may also work.
|
49 |
+
|
50 |
+
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
51 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_InferenceIllusionist__Excalibur-7b-DPO)
|
52 |
+
|
53 |
+
| Metric |Value|
|
54 |
+
|---------------------------------|----:|
|
55 |
+
|Avg. |73.84|
|
56 |
+
|AI2 Reasoning Challenge (25-Shot)|70.90|
|
57 |
+
|HellaSwag (10-Shot) |87.93|
|
58 |
+
|MMLU (5-Shot) |65.46|
|
59 |
+
|TruthfulQA (0-shot) |70.82|
|
60 |
+
|Winogrande (5-shot) |82.48|
|
61 |
+
|GSM8k (5-shot) |65.43|
|