teknium commited on
Commit
cd9634a
1 Parent(s): bcc15ba

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +158 -0
README.md ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: llama2
3
+ language:
4
+ - en
5
+ tags:
6
+ - llama-2
7
+ - self-instruct
8
+ - distillation
9
+ - synthetic instruction
10
+ ---
11
+
12
+ # Model Card: Nous-Hermes-Llama2-13b
13
+
14
+ Compute provided by our project sponsor Redmond AI, thank you! Follow RedmondAI on Twitter @RedmondAI.
15
+
16
+ ## Model Description
17
+
18
+ Nous-Hermes-Llama2-13b is a state-of-the-art language model fine-tuned on over 300,000 instructions. This model was fine-tuned by Nous Research, with Teknium and Emozilla leading the fine tuning process and dataset curation, Redmond AI sponsoring the compute, and several other contributors.
19
+
20
+ This Hermes model uses the exact same dataset as Hermes on Llama-1. This is to ensure consistency between the old Hermes and new, for anyone who wanted to keep Hermes as similar to the old one, just more capable.
21
+
22
+ This model stands out for its long responses, lower hallucination rate, and absence of OpenAI censorship mechanisms. The fine-tuning process was performed with a 4096 sequence length on an 8x a100 80GB DGX machine.
23
+
24
+ ## Example Outputs:
25
+ ![Example4](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example5.png "Example 4")
26
+ ![Example1](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/Example1.png "Example 1")
27
+ ![Example2](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example2.png "Example 2")
28
+ ![Example3](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example3.png "Example 3")
29
+
30
+ ## Model Training
31
+
32
+ The model was trained almost entirely on synthetic GPT-4 outputs. Curating high quality GPT-4 datasets enables incredibly high quality in knowledge, task completion, and style.
33
+
34
+ This includes data from diverse sources such as GPTeacher, the general, roleplay v1&2, code instruct datasets, Nous Instruct & PDACTL (unpublished), and several others, detailed further below
35
+
36
+ ## Collaborators
37
+ The model fine-tuning and the datasets were a collaboration of efforts and resources between Teknium, Karan4D, Emozilla, Huemin Art, and Redmond AI.
38
+
39
+ Special mention goes to @winglian for assisting in some of the training issues.
40
+
41
+ Huge shoutout and acknowledgement is deserved for all the dataset creators who generously share their datasets openly.
42
+
43
+ Among the contributors of datasets:
44
+ - GPTeacher was made available by Teknium
45
+ - Wizard LM by nlpxucan
46
+ - Nous Research Instruct Dataset was provided by Karan4D and HueminArt.
47
+ - GPT4-LLM and Unnatural Instructions were provided by Microsoft
48
+ - Airoboros dataset by jondurbin
49
+ - Camel-AI's domain expert datasets are from Camel-AI
50
+ - CodeAlpaca dataset by Sahil 2801.
51
+
52
+ If anyone was left out, please open a thread in the community tab.
53
+
54
+ ## Prompt Format
55
+
56
+ The model follows the Alpaca prompt format:
57
+ ```
58
+ ### Instruction:
59
+ <prompt>
60
+
61
+ ### Response:
62
+ <leave a newline blank for model to respond>
63
+
64
+ ```
65
+
66
+ or
67
+
68
+ ```
69
+ ### Instruction:
70
+ <prompt>
71
+
72
+ ### Input:
73
+ <additional context>
74
+
75
+ ### Response:
76
+ <leave a newline blank for model to respond>
77
+
78
+ ```
79
+
80
+ ## Benchmark Results
81
+ AGI-Eval
82
+ ```
83
+ | Task |Version| Metric |Value | |Stderr|
84
+ |agieval_aqua_rat | 0|acc |0.2362|± |0.0267|
85
+ | | |acc_norm|0.2480|± |0.0272|
86
+ |agieval_logiqa_en | 0|acc |0.3425|± |0.0186|
87
+ | | |acc_norm|0.3472|± |0.0187|
88
+ |agieval_lsat_ar | 0|acc |0.2522|± |0.0287|
89
+ | | |acc_norm|0.2087|± |0.0269|
90
+ |agieval_lsat_lr | 0|acc |0.3510|± |0.0212|
91
+ | | |acc_norm|0.3627|± |0.0213|
92
+ |agieval_lsat_rc | 0|acc |0.4647|± |0.0305|
93
+ | | |acc_norm|0.4424|± |0.0303|
94
+ |agieval_sat_en | 0|acc |0.6602|± |0.0331|
95
+ | | |acc_norm|0.6165|± |0.0340|
96
+ |agieval_sat_en_without_passage| 0|acc |0.4320|± |0.0346|
97
+ | | |acc_norm|0.4272|± |0.0345|
98
+ |agieval_sat_math | 0|acc |0.2909|± |0.0307|
99
+ | | |acc_norm|0.2727|± |0.0301|
100
+ ```
101
+ GPT-4All Benchmark Set
102
+ ```
103
+ | Task |Version| Metric |Value | |Stderr|
104
+ |arc_challenge| 0|acc |0.5102|± |0.0146|
105
+ | | |acc_norm|0.5213|± |0.0146|
106
+ |arc_easy | 0|acc |0.7959|± |0.0083|
107
+ | | |acc_norm|0.7567|± |0.0088|
108
+ |boolq | 1|acc |0.8394|± |0.0064|
109
+ |hellaswag | 0|acc |0.6164|± |0.0049|
110
+ | | |acc_norm|0.8009|± |0.0040|
111
+ |openbookqa | 0|acc |0.3580|± |0.0215|
112
+ | | |acc_norm|0.4620|± |0.0223|
113
+ |piqa | 0|acc |0.7992|± |0.0093|
114
+ | | |acc_norm|0.8069|± |0.0092|
115
+ |winogrande | 0|acc |0.7127|± |0.0127|
116
+ ```
117
+ BigBench Reasoning Test
118
+ ```
119
+ | Task |Version| Metric |Value | |Stderr|
120
+
121
+ |bigbench_causal_judgement | 0|multiple_choice_grade|0.5526|± |0.0362|
122
+ |bigbench_date_understanding | 0|multiple_choice_grade|0.7344|± |0.0230|
123
+ |bigbench_disambiguation_qa | 0|multiple_choice_grade|0.2636|± |0.0275|
124
+ |bigbench_geometric_shapes | 0|multiple_choice_grade|0.0195|± |0.0073|
125
+ | | |exact_str_match |0.0000|± |0.0000|
126
+ |bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|0.2760|± |0.0200|
127
+ |bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|0.2100|± |0.0154|
128
+ |bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|0.4400|± |0.0287|
129
+ |bigbench_movie_recommendation | 0|multiple_choice_grade|0.2440|± |0.0192|
130
+ |bigbench_navigate | 0|multiple_choice_grade|0.4950|± |0.0158|
131
+ |bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|0.5570|± |0.0111|
132
+ |bigbench_ruin_names | 0|multiple_choice_grade|0.3728|± |0.0229|
133
+ |bigbench_salient_translation_error_detection | 0|multiple_choice_grade|0.1854|± |0.0123|
134
+ |bigbench_snarks | 0|multiple_choice_grade|0.6298|± |0.0360|
135
+ |bigbench_sports_understanding | 0|multiple_choice_grade|0.6156|± |0.0155|
136
+ |bigbench_temporal_sequences | 0|multiple_choice_grade|0.3140|± |0.0147|
137
+ |bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|0.2032|± |0.0114|
138
+ |bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|0.1406|± |0.0083|
139
+ |bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|0.4400|± |0.0287|
140
+ ```
141
+
142
+ These are the highest benchmarks Hermes has seen on every metric, achieving the following average scores:
143
+ - GPT4All benchmark average is now 70.0 - from 68.8 in Hermes-Llama1
144
+ - 0.3657 on BigBench, up from 0.328 on hermes-llama1
145
+ - 0.372 on AGIEval, up from 0.354 on Hermes-llama1
146
+
147
+ These benchmarks currently have us at #1 on ARC-c, ARC-e, Hellaswag, and OpenBookQA, and 2nd place on Winogrande, comparing to GPT4all's benchmarking list, supplanting Hermes 1 for the new top position.
148
+
149
+ ## Resources for Applied Use Cases:
150
+ For an example of a back and forth chatbot using huggingface transformers and discord, check out: https://github.com/teknium1/alpaca-discord
151
+ For an example of a roleplaying discord chatbot, check out this: https://github.com/teknium1/alpaca-roleplay-discordbot
152
+
153
+ ## Future Plans
154
+ We plan to continue to iterate on both more high quality data, and new data filtering techniques to eliminate lower quality data going forward.
155
+
156
+ ## Model Usage
157
+ The model is available for download on Hugging Face. It is suitable for a wide range of language tasks, from generating creative text to understanding and following complex instructions.
158
+