Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
inference: false
|
4 |
+
---
|
5 |
+
|
6 |
+
# Airoboros 13B HF fp16
|
7 |
+
|
8 |
+
These files are HF format fp16 model files for [Jon Durbin's Airoboros 13B](https://huggingface.co/jondurbin/airoboros-13b).
|
9 |
+
|
10 |
+
It is the result of converting Jon's fp32 repo to fp16 for easier storage and usage.
|
11 |
+
|
12 |
+
## Other repositories available
|
13 |
+
|
14 |
+
* [4bit GPTQ models for GPU inference](https://huggingface.co/TheBloke/airoboros-13B-GPTQ)
|
15 |
+
* [Unquantised model in HF fp16 format](https://huggingface.co/TheBloke/airoboros-13B-HF)
|
16 |
+
* [latimar's GGML models for CPU (+CUDA) inference](https://huggingface.co/latimar/airoboros-13b-ggml)
|
17 |
+
|
18 |
+
# Airoboros-13B original model card
|
19 |
+
|
20 |
+
# Overview
|
21 |
+
|
22 |
+
This is a fine-tuned 13b parameter LlaMa model, using completely synthetic training data created by https://github.com/jondurbin/airoboros
|
23 |
+
|
24 |
+
### Eval (gpt4 judging)
|
25 |
+
|
26 |
+

|
27 |
+
|
28 |
+
| model | raw score | gpt-3.5 adjusted score |
|
29 |
+
| --- | --- | --- |
|
30 |
+
| __airoboros-13b__ | __17947__ | __98.087__ |
|
31 |
+
| gpt35 | 18297 | 100.0 |
|
32 |
+
| gpt4-x-alpasta-30b | 15612 | 85.33 |
|
33 |
+
| manticore-13b | 15856 | 86.66 |
|
34 |
+
| vicuna-13b-1.1 | 16306 | 89.12 |
|
35 |
+
| wizard-vicuna-13b-uncensored | 16287 | 89.01 |
|
36 |
+
|
37 |
+
<details>
|
38 |
+
<summary>individual question scores, with shareGPT links (200 prompts generated by gpt-4)</summary>
|
39 |
+
|
40 |
+
*wb-13b-u is Wizard-Vicuna-13b-Uncensored*
|
41 |
+
|
42 |
+
| airoboros-13b | gpt35 | gpt4-x-alpasta-30b | manticore-13b | vicuna-13b-1.1 | wv-13b-u | link |
|
43 |
+
|----------------:|--------:|---------------------:|----------------:|-----------------:|-------------------------------:|:---------------------------------------|
|
44 |
+
| 80 | 95 | 70 | 90 | 85 | 60 | [eval](https://sharegpt.com/c/PIbRQD3) |
|
45 |
+
| 20 | 95 | 40 | 30 | 90 | 80 | [eval](https://sharegpt.com/c/fSzwzzd) |
|
46 |
+
| 100 | 100 | 100 | 95 | 95 | 100 | [eval](https://sharegpt.com/c/AXMzZiO) |
|
47 |
+
| 90 | 100 | 85 | 60 | 95 | 100 | [eval](https://sharegpt.com/c/7obzJm2) |
|
48 |
+
| 95 | 90 | 80 | 85 | 95 | 75 | [eval](https://sharegpt.com/c/cRpj6M1) |
|
49 |
+
| 100 | 95 | 90 | 95 | 98 | 92 | [eval](https://sharegpt.com/c/p0by1T7) |
|
50 |
+
| 50 | 100 | 80 | 95 | 60 | 55 | [eval](https://sharegpt.com/c/rowNlKx) |
|
51 |
+
| 70 | 90 | 80 | 60 | 85 | 40 | [eval](https://sharegpt.com/c/I4POj4I) |
|
52 |
+
| 100 | 95 | 50 | 85 | 40 | 60 | [eval](https://sharegpt.com/c/gUAeiRp) |
|
53 |
+
| 85 | 60 | 55 | 65 | 50 | 70 | [eval](https://sharegpt.com/c/Lgw4QQL) |
|
54 |
+
| 95 | 100 | 85 | 90 | 60 | 75 | [eval](https://sharegpt.com/c/X9tDYft) |
|
55 |
+
| 100 | 95 | 70 | 80 | 50 | 85 | [eval](https://sharegpt.com/c/9V2ElkH) |
|
56 |
+
| 100 | 95 | 80 | 70 | 60 | 90 | [eval](https://sharegpt.com/c/D5xg6qt) |
|
57 |
+
| 95 | 100 | 70 | 85 | 90 | 90 | [eval](https://sharegpt.com/c/lQnSfDs) |
|
58 |
+
| 80 | 95 | 90 | 60 | 30 | 85 | [eval](https://sharegpt.com/c/1hpHGNc) |
|
59 |
+
| 60 | 95 | 0 | 75 | 50 | 40 | [eval](https://sharegpt.com/c/an6TqE4) |
|
60 |
+
| 100 | 95 | 90 | 98 | 95 | 95 | [eval](https://sharegpt.com/c/7vr6n3F) |
|
61 |
+
| 60 | 85 | 40 | 50 | 20 | 0 | [eval](https://sharegpt.com/c/TOkMkgE) |
|
62 |
+
| 100 | 90 | 85 | 95 | 95 | 80 | [eval](https://sharegpt.com/c/Qu7ak0r) |
|
63 |
+
| 100 | 95 | 100 | 95 | 90 | 95 | [eval](https://sharegpt.com/c/hMD4gPo) |
|
64 |
+
| 95 | 90 | 96 | 80 | 92 | 88 | [eval](https://sharegpt.com/c/HTlicNh) |
|
65 |
+
| 95 | 92 | 90 | 93 | 89 | 91 | [eval](https://sharegpt.com/c/MjxHpAf) |
|
66 |
+
| 95 | 93 | 90 | 94 | 96 | 92 | [eval](https://sharegpt.com/c/4RvxOR9) |
|
67 |
+
| 95 | 90 | 93 | 88 | 92 | 85 | [eval](https://sharegpt.com/c/PcAIU9r) |
|
68 |
+
| 95 | 90 | 85 | 96 | 88 | 92 | [eval](https://sharegpt.com/c/MMqul3q) |
|
69 |
+
| 95 | 95 | 90 | 93 | 92 | 91 | [eval](https://sharegpt.com/c/YQsLyzJ) |
|
70 |
+
| 95 | 98 | 80 | 97 | 99 | 96 | [eval](https://sharegpt.com/c/UDhSTMq) |
|
71 |
+
| 95 | 93 | 90 | 87 | 92 | 89 | [eval](https://sharegpt.com/c/4gCfdCV) |
|
72 |
+
| 90 | 85 | 95 | 80 | 92 | 75 | [eval](https://sharegpt.com/c/bkQs4SP) |
|
73 |
+
| 90 | 85 | 95 | 93 | 80 | 92 | [eval](https://sharegpt.com/c/LeLCEEt) |
|
74 |
+
| 95 | 92 | 90 | 91 | 93 | 89 | [eval](https://sharegpt.com/c/DFxNzVu) |
|
75 |
+
| 100 | 95 | 90 | 85 | 80 | 95 | [eval](https://sharegpt.com/c/gnVzNML) |
|
76 |
+
| 95 | 97 | 93 | 92 | 96 | 94 | [eval](https://sharegpt.com/c/y7pxMIy) |
|
77 |
+
| 95 | 93 | 94 | 90 | 88 | 92 | [eval](https://sharegpt.com/c/5UeCvTY) |
|
78 |
+
| 90 | 95 | 98 | 85 | 96 | 92 | [eval](https://sharegpt.com/c/T4oL9I5) |
|
79 |
+
| 90 | 88 | 85 | 80 | 82 | 84 | [eval](https://sharegpt.com/c/HnGyTAG) |
|
80 |
+
| 90 | 95 | 85 | 87 | 92 | 88 | [eval](https://sharegpt.com/c/ZbRMBNj) |
|
81 |
+
| 95 | 97 | 96 | 90 | 93 | 92 | [eval](https://sharegpt.com/c/iTmFJqd) |
|
82 |
+
| 95 | 93 | 92 | 90 | 89 | 91 | [eval](https://sharegpt.com/c/VuPifET) |
|
83 |
+
| 90 | 95 | 93 | 92 | 94 | 91 | [eval](https://sharegpt.com/c/AvFAH1x) |
|
84 |
+
| 90 | 85 | 95 | 80 | 88 | 75 | [eval](https://sharegpt.com/c/4ealKGN) |
|
85 |
+
| 85 | 90 | 95 | 88 | 92 | 80 | [eval](https://sharegpt.com/c/bE1b2vX) |
|
86 |
+
| 90 | 95 | 92 | 85 | 80 | 87 | [eval](https://sharegpt.com/c/I3nMPBC) |
|
87 |
+
| 85 | 90 | 95 | 80 | 88 | 75 | [eval](https://sharegpt.com/c/as7r3bW) |
|
88 |
+
| 85 | 80 | 75 | 90 | 70 | 82 | [eval](https://sharegpt.com/c/qYceaUa) |
|
89 |
+
| 90 | 85 | 95 | 92 | 93 | 80 | [eval](https://sharegpt.com/c/g4FXchU) |
|
90 |
+
| 90 | 95 | 75 | 85 | 80 | 70 | [eval](https://sharegpt.com/c/6kGLvL5) |
|
91 |
+
| 85 | 90 | 80 | 88 | 82 | 83 | [eval](https://sharegpt.com/c/SRozqaF) |
|
92 |
+
| 85 | 90 | 95 | 92 | 88 | 80 | [eval](https://sharegpt.com/c/GoKydf6) |
|
93 |
+
| 85 | 90 | 80 | 75 | 95 | 88 | [eval](https://sharegpt.com/c/37aXkHQ) |
|
94 |
+
| 85 | 90 | 80 | 88 | 84 | 92 | [eval](https://sharegpt.com/c/nVuUaTj) |
|
95 |
+
| 80 | 90 | 75 | 85 | 70 | 95 | [eval](https://sharegpt.com/c/TkAQKLC) |
|
96 |
+
| 90 | 88 | 85 | 80 | 92 | 83 | [eval](https://sharegpt.com/c/55cO2y0) |
|
97 |
+
| 85 | 75 | 90 | 80 | 78 | 88 | [eval](https://sharegpt.com/c/tXtq5lT) |
|
98 |
+
| 85 | 90 | 80 | 82 | 75 | 88 | [eval](https://sharegpt.com/c/TfMjeJQ) |
|
99 |
+
| 90 | 85 | 40 | 95 | 80 | 88 | [eval](https://sharegpt.com/c/2jQ6K2S) |
|
100 |
+
| 85 | 95 | 90 | 75 | 88 | 80 | [eval](https://sharegpt.com/c/aQtr2ca) |
|
101 |
+
| 85 | 95 | 90 | 92 | 89 | 88 | [eval](https://sharegpt.com/c/tbWLyZ7) |
|
102 |
+
| 80 | 85 | 75 | 60 | 90 | 70 | [eval](https://sharegpt.com/c/moHC7i2) |
|
103 |
+
| 85 | 90 | 87 | 80 | 88 | 75 | [eval](https://sharegpt.com/c/GK6GShh) |
|
104 |
+
| 85 | 80 | 75 | 50 | 90 | 80 | [eval](https://sharegpt.com/c/ugcW4qG) |
|
105 |
+
| 95 | 80 | 90 | 85 | 75 | 82 | [eval](https://sharegpt.com/c/WL8iq6F) |
|
106 |
+
| 85 | 90 | 80 | 70 | 95 | 88 | [eval](https://sharegpt.com/c/TZJKnvS) |
|
107 |
+
| 90 | 95 | 70 | 85 | 80 | 75 | [eval](https://sharegpt.com/c/beNOKb5) |
|
108 |
+
| 90 | 85 | 70 | 75 | 80 | 60 | [eval](https://sharegpt.com/c/o2oRCF5) |
|
109 |
+
| 95 | 90 | 70 | 50 | 85 | 80 | [eval](https://sharegpt.com/c/TNjbK6D) |
|
110 |
+
| 80 | 85 | 40 | 60 | 90 | 95 | [eval](https://sharegpt.com/c/rJvszWJ) |
|
111 |
+
| 75 | 60 | 80 | 55 | 70 | 85 | [eval](https://sharegpt.com/c/HJwRkro) |
|
112 |
+
| 90 | 85 | 60 | 50 | 80 | 95 | [eval](https://sharegpt.com/c/AeFoSDK) |
|
113 |
+
| 45 | 85 | 60 | 20 | 65 | 75 | [eval](https://sharegpt.com/c/KA1cgOl) |
|
114 |
+
| 85 | 90 | 30 | 60 | 80 | 70 | [eval](https://sharegpt.com/c/RTy8n0y) |
|
115 |
+
| 90 | 95 | 80 | 40 | 85 | 70 | [eval](https://sharegpt.com/c/PJMJoXh) |
|
116 |
+
| 85 | 90 | 70 | 75 | 80 | 95 | [eval](https://sharegpt.com/c/Ib3jzyC) |
|
117 |
+
| 90 | 70 | 50 | 20 | 60 | 40 | [eval](https://sharegpt.com/c/oMmqqtX) |
|
118 |
+
| 90 | 95 | 75 | 60 | 85 | 80 | [eval](https://sharegpt.com/c/qRNhNTw) |
|
119 |
+
| 85 | 80 | 60 | 70 | 65 | 75 | [eval](https://sharegpt.com/c/3MAHQIy) |
|
120 |
+
| 90 | 85 | 80 | 75 | 82 | 70 | [eval](https://sharegpt.com/c/0Emc5HS) |
|
121 |
+
| 90 | 95 | 80 | 70 | 85 | 75 | [eval](https://sharegpt.com/c/UqAxRWF) |
|
122 |
+
| 85 | 75 | 30 | 80 | 90 | 70 | [eval](https://sharegpt.com/c/eywxGAw) |
|
123 |
+
| 85 | 90 | 50 | 70 | 80 | 60 | [eval](https://sharegpt.com/c/A2KSEWP) |
|
124 |
+
| 100 | 95 | 98 | 99 | 97 | 96 | [eval](https://sharegpt.com/c/C8rebQf) |
|
125 |
+
| 95 | 90 | 92 | 93 | 91 | 89 | [eval](https://sharegpt.com/c/cd9HF4V) |
|
126 |
+
| 95 | 92 | 90 | 85 | 88 | 91 | [eval](https://sharegpt.com/c/LHkjvQJ) |
|
127 |
+
| 100 | 95 | 98 | 97 | 96 | 99 | [eval](https://sharegpt.com/c/o5PdoyZ) |
|
128 |
+
| 100 | 100 | 100 | 90 | 100 | 95 | [eval](https://sharegpt.com/c/rh8pZVg) |
|
129 |
+
| 100 | 95 | 98 | 97 | 94 | 99 | [eval](https://sharegpt.com/c/T5DYL83) |
|
130 |
+
| 95 | 90 | 92 | 93 | 94 | 91 | [eval](https://sharegpt.com/c/G5Osg3X) |
|
131 |
+
| 100 | 95 | 98 | 90 | 96 | 95 | [eval](https://sharegpt.com/c/9ZqI03V) |
|
132 |
+
| 95 | 96 | 92 | 90 | 89 | 93 | [eval](https://sharegpt.com/c/4tFfwZU) |
|
133 |
+
| 100 | 95 | 93 | 90 | 92 | 88 | [eval](https://sharegpt.com/c/mG1JqPH) |
|
134 |
+
| 100 | 100 | 98 | 97 | 99 | 100 | [eval](https://sharegpt.com/c/VDdtgCu) |
|
135 |
+
| 95 | 90 | 92 | 85 | 93 | 94 | [eval](https://sharegpt.com/c/uKtGkvg) |
|
136 |
+
| 95 | 93 | 90 | 92 | 96 | 91 | [eval](https://sharegpt.com/c/9B92N6P) |
|
137 |
+
| 95 | 96 | 92 | 90 | 93 | 91 | [eval](https://sharegpt.com/c/GeIFfOu) |
|
138 |
+
| 95 | 90 | 92 | 93 | 91 | 89 | [eval](https://sharegpt.com/c/gn3E9nN) |
|
139 |
+
| 100 | 98 | 95 | 97 | 96 | 99 | [eval](https://sharegpt.com/c/Erxa46H) |
|
140 |
+
| 90 | 95 | 85 | 88 | 92 | 87 | [eval](https://sharegpt.com/c/oRHVOvK) |
|
141 |
+
| 95 | 93 | 90 | 92 | 89 | 88 | [eval](https://sharegpt.com/c/ghtKLUX) |
|
142 |
+
| 100 | 95 | 97 | 90 | 96 | 94 | [eval](https://sharegpt.com/c/ZL4KjqP) |
|
143 |
+
| 95 | 93 | 90 | 92 | 94 | 91 | [eval](https://sharegpt.com/c/YOnqIQa) |
|
144 |
+
| 95 | 92 | 90 | 93 | 94 | 88 | [eval](https://sharegpt.com/c/3BKwKho) |
|
145 |
+
| 95 | 92 | 60 | 97 | 90 | 96 | [eval](https://sharegpt.com/c/U1i31bn) |
|
146 |
+
| 95 | 90 | 92 | 93 | 91 | 89 | [eval](https://sharegpt.com/c/etfRoAE) |
|
147 |
+
| 95 | 90 | 97 | 92 | 91 | 93 | [eval](https://sharegpt.com/c/B0OpVxR) |
|
148 |
+
| 90 | 95 | 93 | 85 | 92 | 91 | [eval](https://sharegpt.com/c/MBgGJ5A) |
|
149 |
+
| 95 | 90 | 40 | 92 | 93 | 85 | [eval](https://sharegpt.com/c/eQKTYO7) |
|
150 |
+
| 100 | 100 | 95 | 90 | 95 | 90 | [eval](https://sharegpt.com/c/szKWCBt) |
|
151 |
+
| 90 | 95 | 96 | 98 | 93 | 92 | [eval](https://sharegpt.com/c/8ZhUcAv) |
|
152 |
+
| 90 | 95 | 92 | 89 | 93 | 94 | [eval](https://sharegpt.com/c/VQWdy99) |
|
153 |
+
| 100 | 95 | 100 | 98 | 96 | 99 | [eval](https://sharegpt.com/c/g1DHUSM) |
|
154 |
+
| 100 | 100 | 95 | 90 | 100 | 90 | [eval](https://sharegpt.com/c/uYgfJC3) |
|
155 |
+
| 90 | 85 | 88 | 92 | 87 | 91 | [eval](https://sharegpt.com/c/crk8BH3) |
|
156 |
+
| 95 | 97 | 90 | 92 | 93 | 94 | [eval](https://sharegpt.com/c/95F9afQ) |
|
157 |
+
| 90 | 95 | 85 | 88 | 92 | 89 | [eval](https://sharegpt.com/c/otioHUo) |
|
158 |
+
| 95 | 93 | 90 | 92 | 94 | 91 | [eval](https://sharegpt.com/c/KSiL9F6) |
|
159 |
+
| 90 | 95 | 85 | 80 | 88 | 82 | [eval](https://sharegpt.com/c/GmGq3b3) |
|
160 |
+
| 95 | 90 | 60 | 85 | 93 | 70 | [eval](https://sharegpt.com/c/VOhklyz) |
|
161 |
+
| 95 | 92 | 94 | 93 | 96 | 90 | [eval](https://sharegpt.com/c/wqy8m6k) |
|
162 |
+
| 95 | 90 | 85 | 93 | 87 | 92 | [eval](https://sharegpt.com/c/iWKrIuS) |
|
163 |
+
| 95 | 96 | 93 | 90 | 97 | 92 | [eval](https://sharegpt.com/c/o1h3w8N) |
|
164 |
+
| 100 | 0 | 0 | 100 | 0 | 0 | [eval](https://sharegpt.com/c/3UH9eed) |
|
165 |
+
| 60 | 100 | 0 | 80 | 0 | 0 | [eval](https://sharegpt.com/c/44g0FAh) |
|
166 |
+
| 0 | 100 | 60 | 0 | 0 | 90 | [eval](https://sharegpt.com/c/PaQlcrU) |
|
167 |
+
| 100 | 100 | 0 | 100 | 100 | 100 | [eval](https://sharegpt.com/c/51icV4o) |
|
168 |
+
| 100 | 100 | 100 | 100 | 95 | 100 | [eval](https://sharegpt.com/c/1VnbGAR) |
|
169 |
+
| 100 | 100 | 100 | 50 | 90 | 100 | [eval](https://sharegpt.com/c/EYGBrgw) |
|
170 |
+
| 100 | 100 | 100 | 100 | 95 | 90 | [eval](https://sharegpt.com/c/EGRduOt) |
|
171 |
+
| 100 | 100 | 100 | 95 | 0 | 100 | [eval](https://sharegpt.com/c/O3JJfnK) |
|
172 |
+
| 50 | 95 | 20 | 10 | 30 | 85 | [eval](https://sharegpt.com/c/2roVtAu) |
|
173 |
+
| 100 | 100 | 60 | 20 | 30 | 40 | [eval](https://sharegpt.com/c/sphFpfx) |
|
174 |
+
| 100 | 0 | 0 | 0 | 0 | 100 | [eval](https://sharegpt.com/c/OeWGKBo) |
|
175 |
+
| 0 | 100 | 60 | 0 | 0 | 80 | [eval](https://sharegpt.com/c/TOUsuFA) |
|
176 |
+
| 50 | 100 | 20 | 90 | 0 | 10 | [eval](https://sharegpt.com/c/Y3P6DCu) |
|
177 |
+
| 100 | 100 | 100 | 100 | 100 | 100 | [eval](https://sharegpt.com/c/hkbdeiM) |
|
178 |
+
| 100 | 100 | 100 | 100 | 100 | 100 | [eval](https://sharegpt.com/c/eubbaVC) |
|
179 |
+
| 40 | 100 | 95 | 0 | 100 | 40 | [eval](https://sharegpt.com/c/QWiF49v) |
|
180 |
+
| 100 | 100 | 100 | 100 | 80 | 100 | [eval](https://sharegpt.com/c/dKTapBu) |
|
181 |
+
| 100 | 100 | 100 | 0 | 90 | 40 | [eval](https://sharegpt.com/c/P8NGwFZ) |
|
182 |
+
| 0 | 100 | 100 | 50 | 70 | 20 | [eval](https://sharegpt.com/c/v96BtBL) |
|
183 |
+
| 100 | 100 | 50 | 90 | 0 | 95 | [eval](https://sharegpt.com/c/YRlzj1t) |
|
184 |
+
| 100 | 95 | 90 | 85 | 98 | 80 | [eval](https://sharegpt.com/c/76VX3eB) |
|
185 |
+
| 95 | 98 | 90 | 92 | 96 | 89 | [eval](https://sharegpt.com/c/JK1uNef) |
|
186 |
+
| 90 | 95 | 75 | 85 | 80 | 82 | [eval](https://sharegpt.com/c/ku6CKmx) |
|
187 |
+
| 95 | 98 | 50 | 92 | 96 | 94 | [eval](https://sharegpt.com/c/0iAFuKW) |
|
188 |
+
| 95 | 90 | 0 | 93 | 92 | 94 | [eval](https://sharegpt.com/c/6uGnKio) |
|
189 |
+
| 95 | 90 | 85 | 92 | 80 | 88 | [eval](https://sharegpt.com/c/lfpRBw8) |
|
190 |
+
| 95 | 93 | 75 | 85 | 90 | 92 | [eval](https://sharegpt.com/c/mKu70jb) |
|
191 |
+
| 90 | 95 | 88 | 85 | 92 | 89 | [eval](https://sharegpt.com/c/GkYzJHO) |
|
192 |
+
| 100 | 100 | 100 | 95 | 97 | 98 | [eval](https://sharegpt.com/c/mly2k0z) |
|
193 |
+
| 85 | 40 | 30 | 95 | 90 | 88 | [eval](https://sharegpt.com/c/5td2ob0) |
|
194 |
+
| 90 | 95 | 92 | 85 | 88 | 93 | [eval](https://sharegpt.com/c/0ISpWfy) |
|
195 |
+
| 95 | 96 | 92 | 90 | 89 | 93 | [eval](https://sharegpt.com/c/kdUDUn7) |
|
196 |
+
| 90 | 95 | 85 | 80 | 92 | 88 | [eval](https://sharegpt.com/c/fjMNYr2) |
|
197 |
+
| 95 | 98 | 65 | 90 | 85 | 93 | [eval](https://sharegpt.com/c/6xBIf2Q) |
|
198 |
+
| 95 | 92 | 96 | 97 | 90 | 89 | [eval](https://sharegpt.com/c/B9GY8Ln) |
|
199 |
+
| 95 | 90 | 92 | 91 | 89 | 93 | [eval](https://sharegpt.com/c/vn1FPU4) |
|
200 |
+
| 95 | 90 | 80 | 75 | 95 | 90 | [eval](https://sharegpt.com/c/YurEMYg) |
|
201 |
+
| 92 | 40 | 30 | 95 | 90 | 93 | [eval](https://sharegpt.com/c/D19Qeui) |
|
202 |
+
| 90 | 92 | 85 | 88 | 89 | 87 | [eval](https://sharegpt.com/c/5QRFfrt) |
|
203 |
+
| 95 | 80 | 90 | 92 | 91 | 88 | [eval](https://sharegpt.com/c/pYWPRi4) |
|
204 |
+
| 95 | 93 | 92 | 90 | 91 | 94 | [eval](https://sharegpt.com/c/wPRTntL) |
|
205 |
+
| 100 | 98 | 95 | 90 | 92 | 96 | [eval](https://sharegpt.com/c/F6PLYKE) |
|
206 |
+
| 95 | 92 | 80 | 85 | 90 | 93 | [eval](https://sharegpt.com/c/WeJnMGv) |
|
207 |
+
| 95 | 98 | 90 | 88 | 97 | 96 | [eval](https://sharegpt.com/c/zNKL49e) |
|
208 |
+
| 90 | 95 | 85 | 88 | 86 | 92 | [eval](https://sharegpt.com/c/kIKmA1b) |
|
209 |
+
| 100 | 100 | 100 | 100 | 100 | 100 | [eval](https://sharegpt.com/c/1btWd4O) |
|
210 |
+
| 90 | 95 | 85 | 96 | 92 | 88 | [eval](https://sharegpt.com/c/s9sf1Lp) |
|
211 |
+
| 100 | 98 | 95 | 99 | 97 | 96 | [eval](https://sharegpt.com/c/RWzv8py) |
|
212 |
+
| 95 | 92 | 70 | 90 | 93 | 89 | [eval](https://sharegpt.com/c/bYF7FqA) |
|
213 |
+
| 95 | 90 | 88 | 92 | 94 | 93 | [eval](https://sharegpt.com/c/SuUqjMj) |
|
214 |
+
| 95 | 90 | 93 | 92 | 85 | 94 | [eval](https://sharegpt.com/c/r0aRdYY) |
|
215 |
+
| 95 | 93 | 90 | 87 | 92 | 91 | [eval](https://sharegpt.com/c/VuMfkkd) |
|
216 |
+
| 95 | 93 | 90 | 96 | 92 | 91 | [eval](https://sharegpt.com/c/rhm6fa4) |
|
217 |
+
| 95 | 97 | 85 | 96 | 98 | 90 | [eval](https://sharegpt.com/c/DwXnyqG) |
|
218 |
+
| 95 | 92 | 90 | 85 | 93 | 94 | [eval](https://sharegpt.com/c/0ScdkGS) |
|
219 |
+
| 95 | 96 | 92 | 90 | 97 | 93 | [eval](https://sharegpt.com/c/6yIoCDU) |
|
220 |
+
| 95 | 93 | 96 | 94 | 90 | 92 | [eval](https://sharegpt.com/c/VubEvp9) |
|
221 |
+
| 95 | 94 | 93 | 92 | 90 | 89 | [eval](https://sharegpt.com/c/RHzmZWG) |
|
222 |
+
| 90 | 85 | 95 | 80 | 87 | 75 | [eval](https://sharegpt.com/c/IMiP9Zm) |
|
223 |
+
| 95 | 94 | 92 | 93 | 90 | 96 | [eval](https://sharegpt.com/c/bft4PIL) |
|
224 |
+
| 95 | 100 | 90 | 95 | 95 | 95 | [eval](https://sharegpt.com/c/iHXB34b) |
|
225 |
+
| 100 | 95 | 85 | 100 | 0 | 90 | [eval](https://sharegpt.com/c/vCGn9R7) |
|
226 |
+
| 100 | 95 | 90 | 95 | 100 | 95 | [eval](https://sharegpt.com/c/be8crZL) |
|
227 |
+
| 95 | 90 | 60 | 95 | 85 | 80 | [eval](https://sharegpt.com/c/33elmDz) |
|
228 |
+
| 100 | 95 | 90 | 98 | 97 | 99 | [eval](https://sharegpt.com/c/RWD3Zx7) |
|
229 |
+
| 95 | 90 | 85 | 95 | 80 | 92 | [eval](https://sharegpt.com/c/GiwBvM7) |
|
230 |
+
| 100 | 95 | 100 | 98 | 100 | 90 | [eval](https://sharegpt.com/c/hX2pYxk) |
|
231 |
+
| 100 | 95 | 80 | 85 | 90 | 85 | [eval](https://sharegpt.com/c/MfxdGd7) |
|
232 |
+
| 100 | 90 | 95 | 85 | 95 | 100 | [eval](https://sharegpt.com/c/28hQjmS) |
|
233 |
+
| 95 | 90 | 85 | 80 | 88 | 92 | [eval](https://sharegpt.com/c/fzy5EPe) |
|
234 |
+
| 100 | 100 | 0 | 0 | 100 | 0 | [eval](https://sharegpt.com/c/vwxPjbR) |
|
235 |
+
| 100 | 100 | 100 | 50 | 100 | 75 | [eval](https://sharegpt.com/c/FAYfFWy) |
|
236 |
+
| 100 | 100 | 0 | 0 | 100 | 0 | [eval](https://sharegpt.com/c/SoudGsQ) |
|
237 |
+
| 0 | 100 | 0 | 0 | 0 | 0 | [eval](https://sharegpt.com/c/mkwEgVn) |
|
238 |
+
| 100 | 100 | 50 | 0 | 0 | 0 | [eval](https://sharegpt.com/c/q8MQEsz) |
|
239 |
+
| 100 | 100 | 100 | 100 | 100 | 95 | [eval](https://sharegpt.com/c/tzHpsKh) |
|
240 |
+
| 100 | 100 | 50 | 0 | 0 | 0 | [eval](https://sharegpt.com/c/3ugYBtJ) |
|
241 |
+
| 100 | 100 | 0 | 0 | 100 | 0 | [eval](https://sharegpt.com/c/I6KfOJT) |
|
242 |
+
| 90 | 85 | 80 | 95 | 70 | 75 | [eval](https://sharegpt.com/c/enaV1CK) |
|
243 |
+
| 100 | 100 | 0 | 0 | 0 | 0 | [eval](https://sharegpt.com/c/JBk7oSh) |
|
244 |
+
|
245 |
+
</details>
|
246 |
+
|
247 |
+
|
248 |
+
### Training data
|
249 |
+
|
250 |
+
I used a jailbreak prompt to generate the synthetic instructions, which resulted in some training data that would likely be censored by other models, such as how-to prompts about synthesizing drugs, making homemade flamethrowers, etc. Mind you, this is all generated by ChatGPT, not me. My goal was to simply test some of the capabilities of ChatGPT when unfiltered (as much as possible), and not to intentionally produce any harmful/dangerous/etc. content.
|
251 |
+
|
252 |
+
The jailbreak prompt I used is the default prompt in the python code when using the `--uncensored` flag: https://github.com/jondurbin/airoboros/blob/main/airoboros/self_instruct.py#L39
|
253 |
+
|
254 |
+
I also did a few passes of manually cleanup to remove some bad prompts, but mostly I left the data as-is. Initially, the model was fairly bad at math/extrapolation, closed question-answering (heavy hallucination), and coding, so I did one more fine tuning pass with additional synthetic instructions aimed at those types of problems.
|
255 |
+
|
256 |
+
Both the initial instructions and final-pass fine-tuning instructions will be published soon.
|
257 |
+
|
258 |
+
### Fine-tuning method
|
259 |
+
|
260 |
+
I used the excellent [FastChat](https://github.com/lm-sys/FastChat) module, running with:
|
261 |
+
|
262 |
+
```
|
263 |
+
source /workspace/venv/bin/activate
|
264 |
+
|
265 |
+
export NCCL_P2P_DISABLE=1
|
266 |
+
export NCCL_P2P_LEVEL=LOC
|
267 |
+
|
268 |
+
torchrun --nproc_per_node=8 --master_port=20001 /workspace/FastChat/fastchat/train/train_mem.py \
|
269 |
+
--model_name_or_path /workspace/llama-13b \
|
270 |
+
--data_path /workspace/as_conversations.json \
|
271 |
+
--bf16 True \
|
272 |
+
--output_dir /workspace/airoboros-uncensored-13b \
|
273 |
+
--num_train_epochs 3 \
|
274 |
+
--per_device_train_batch_size 20 \
|
275 |
+
--per_device_eval_batch_size 20 \
|
276 |
+
--gradient_accumulation_steps 2 \
|
277 |
+
--evaluation_strategy "steps" \
|
278 |
+
--eval_steps 500 \
|
279 |
+
--save_strategy "steps" \
|
280 |
+
--save_steps 500 \
|
281 |
+
--save_total_limit 10 \
|
282 |
+
--learning_rate 2e-5 \
|
283 |
+
--weight_decay 0. \
|
284 |
+
--warmup_ratio 0.04 \
|
285 |
+
--lr_scheduler_type "cosine" \
|
286 |
+
--logging_steps 1 \
|
287 |
+
--fsdp "full_shard auto_wrap offload" \
|
288 |
+
--fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
|
289 |
+
--tf32 True \
|
290 |
+
--model_max_length 2048 \
|
291 |
+
--gradient_checkpointing True \
|
292 |
+
--lazy_preprocess True
|
293 |
+
```
|
294 |
+
|
295 |
+
This ran on 8x nvidia 80gb a100's for about 40 hours.
|
296 |
+
|
297 |
+

|
298 |
+
|
299 |
+

|
300 |
+
|
301 |
+
|
302 |
+
### Prompt format
|
303 |
+
|
304 |
+
The prompt should be 1:1 compatible with the FastChat/vicuna format, e.g.:
|
305 |
+
|
306 |
+
With a preamble:
|
307 |
+
```
|
308 |
+
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
|
309 |
+
|
310 |
+
USER: [prompt]
|
311 |
+
<\s>
|
312 |
+
|
313 |
+
ASSISTANT:
|
314 |
+
```
|
315 |
+
|
316 |
+
Or just:
|
317 |
+
```
|
318 |
+
USER: [prompt]
|
319 |
+
<\s>
|
320 |
+
|
321 |
+
ASSISTANT:
|
322 |
+
```
|
323 |
+
|
324 |
+
### License
|
325 |
+
The model is licensed under the LLaMA model, and the dataset is licensed under the terms of OpenAI because it uses ChatGPT. Everything else is free.
|