Update README.md
Browse files
README.md
CHANGED
|
@@ -1,12 +1,73 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-4.0
|
|
|
|
|
|
|
|
|
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
---
|
| 6 |
|
| 7 |
-
|
| 8 |
|
| 9 |
---
|
| 10 |
|
| 11 |
-
|
|
|
|
| 12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-4.0
|
| 3 |
+
datasets:
|
| 4 |
+
- Setiaku/Stheno-v3.4-Instruct
|
| 5 |
+
- Setiaku/Stheno-3.4-Creative-2
|
| 6 |
language:
|
| 7 |
- en
|
| 8 |
---
|
| 9 |
|
| 10 |
+
|
| 11 |
|
| 12 |
---
|
| 13 |
|
| 14 |
+
Thanks to Gargamel for the compute, to train this!
|
| 15 |
+
<br>It took ~120 Hours on Several A100s.
|
| 16 |
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
Llama-3.1-70B-Euryale-v2.2
|
| 20 |
+
|
| 21 |
+
This model has went through a single stage finetuning process, over 2 epochs. Datasets are cleanly seperated in order, and not merged unlike Stheno v3.4 .
|
| 22 |
+
```
|
| 23 |
+
- 1st, over a multi-turn Conversational-Instruct
|
| 24 |
+
- 2nd, over a Creative Writing / Roleplay along with some Creative-based Instruct Datasets.
|
| 25 |
+
- - Dataset consists of a mixture of Human and Claude Data.
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
Prompting Format:
|
| 29 |
+
```
|
| 30 |
+
- Use the L3 Instruct Formatting - Euryale 2.1 Preset Works Well
|
| 31 |
+
- Temperature + min_p as per usual, I recommend 1.2 Temp + 0.2 min_p.
|
| 32 |
+
- Has a different vibe to previous versions. Tinker around.
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
Changes since Euryale v2.1 \[Same Dataset as Stheno 3.4\]
|
| 36 |
+
```
|
| 37 |
+
- Included Multi-turn Conversation-based Instruct Datasets to boost multi-turn coherency. # This is a seperate set, not the ones made by Kalomaze and Nopm, that are used in Magnum. They're completely different data.
|
| 38 |
+
- Replaced Single-Turn Instruct with Better Prompts and Answers by Claude 3.5 Sonnet and Claude 3 Opus.
|
| 39 |
+
- Removed c2 Samples -> Underway of re-filtering and masking to use with custom prefills. TBD
|
| 40 |
+
- Included 55% more Roleplaying Examples based of [Gryphe's](https://huggingface.co/datasets/Gryphe/Sonnet3.5-Charcard-Roleplay) Charcard RP Sets. Further filtered and cleaned on.
|
| 41 |
+
- Included 40% More Creative Writing Examples.
|
| 42 |
+
- Included Datasets Targeting System Prompt Adherence.
|
| 43 |
+
- Included Datasets targeting Reasoning / Spatial Awareness.
|
| 44 |
+
- Filtered for the usual errors, slop and stuff at the end. Some may have slipped through, but I removed nearly all of it.
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
Personal Opinions:
|
| 48 |
+
```
|
| 49 |
+
- Llama 3.1 is... meh. I'm sure you guys in the community have debated over this.
|
| 50 |
+
- Whatever they did to their Instruct overcooked the model. Base is weird.
|
| 51 |
+
- Still, the 70B is pretty nice, though sometimes it bugs out? A swipe / regen always fixes it.
|
| 52 |
+
- May be less 'uncensored' zero-shot due to removal of c2 samples, but it is perfectly fine for roleplaying purposes.
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
Below are some graphs and all for you to observe.
|
| 56 |
+
|
| 57 |
+
---
|
| 58 |
+
|
| 59 |
+
`Turn Distribution # 1 Turn is considered as 1 combined Human/GPT pair in a ShareGPT format. 4 Turns means 1 System Row + 8 Human/GPT rows in total.`
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
`Token Count Histogram # Based on the Llama 3 Tokenizer`
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
---
|
| 68 |
+
|
| 69 |
+
Have a good one.
|
| 70 |
+
|
| 71 |
+
```
|
| 72 |
+
Source Image: https://www.pixiv.net/en/artworks/91689070
|
| 73 |
+
```
|