ausboss commited on
Commit
39988af
1 Parent(s): 9eba2ad

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +71 -0
README.md ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Merge of [SuperHOT-LoRA-prototype](https://huggingface.co/kaiokendev/SuperHOT-LoRA-prototype) and [llama-30b](https://huggingface.co/huggyllama/llama-30b)
2
+
3
+
4
+ Quantization:
5
+ CUDA_VISIBLE_DEVICES=0 python llama.py ausboss/Llama30B-SuperHOT c4 --wbits 4 --true-sequential --groupsize 128 --save_safetensors Llama30B-SuperHOT-4bit-128g.safetensors
6
+
7
+ Make sure to run with 128g and 4bit arguments in ooba or use the kobold fork that allows for 4bit.
8
+
9
+
10
+ # From the SuperHot Page:
11
+
12
+ ## Prototypes for SuperHOT
13
+
14
+ No guarantees for output quality, simply uploading what I have so others can play around with it. Not even sure if the rank in cutoff-8192 is correct (think it should be 10 maybe.. can't remember)
15
+
16
+ All prototypes are extremely early epochs (sub 0.5)
17
+
18
+ ## Model/Training
19
+ All trained with Flash Attention with conversation sequence lengths ranging from 8K to 16K tokens (No Alibi unless otherwise mentioned)
20
+
21
+ All trained on LLaMa 13B 4-bit (no groupsize)
22
+
23
+ (*Personally, I like the 8K cutoff version better, so I would say start with that one*)
24
+
25
+ ## Data
26
+ A combination of various datasets and cleaned logs converted into datasets including but not limited to:
27
+ - Bluemoon Fanbased
28
+ - Roleplaying Guild
29
+ - Community-sourced outputs
30
+ - [Dan's PocketDoc/RUCAIBox-Story-Generation-Alpaca](https://huggingface.co/datasets/PocketDoc/RUCAIBox-Story-Generation-Alpaca)
31
+ - [IlyaGusev/gpt_roleplay_realm](https://huggingface.co/datasets/IlyaGusev/gpt_roleplay_realm)
32
+ - others
33
+
34
+ ## Bias
35
+ SuperHOT is a fiction-focused model. No alignment has been performed on the training data. Be mindful that this model may output harmful, violent, or otherwise problematic content
36
+
37
+ ## Format
38
+ Any format should work with such early checkpoints. However the training data is entirely in the following format:
39
+ ```
40
+ ---
41
+ mode: chat
42
+ characters:
43
+ <char1 name>: <descriptive tags for char1>
44
+ <char2 name>: <descriptive tags for char2>
45
+ summary: <summary of the story thus far or the purpose of the chat> (optional)
46
+ <any other miscellaneous data>
47
+ ---
48
+ <chat history>
49
+ ```
50
+
51
+ By "any other miscellaneous data", it means you should be able to put any additional metadata for the story or characters. I.e.,
52
+ ```
53
+ ...
54
+ locations:
55
+ location1: <tags for location1>
56
+ inventory:
57
+ item1: <tags for item1>
58
+ ```
59
+
60
+ Again, format does not hold such a large weight on these early checkpoints. I have found success with the following setup for an RPG-like experience. Just play around with the format and see what works:
61
+ ```
62
+ ---
63
+ mode: rpg
64
+ characters:
65
+ You: a new player
66
+ system: The system controls the RPG, handles character creation, world narration, and quest management. Also controls any NPCs and inventory tracking. Their first message provides a lengthy introduction for the player into the RPG world they are about to play in. After completing the character creation, the system will give a lengthy introduction into the world of ___. The first quest will follow right after
67
+ rpg setting: The world of ___
68
+ rpg rules: Any rules typical of RPG games, including typical items, battle stats, etc
69
+ ---
70
+ ```
71
+