Hugging Face's logo

asifahmed
/
open_llama_13b_NH

Text Generation
Transformers
PyTorch
English
llama
llama-2
self-instruct
distillation
synthetic instruction
text-generation-inference
Inference Endpoints
Model card Files Community
Asif Ahmad commited on
Commit
7a2b385
·
1 Parent(s): b84ddcd

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +160 -0
README.md ADDED
@@ -0,0 +1,160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ tags:
5
+ - llama-2
6
+ - self-instruct
7
+ - distillation
8
+ - synthetic instruction
9
+ license:
10
+ - mit
11
+ ---
12
+
13
+ # Model Card: Nous-Hermes-Llama2-13b
14
+
15
+ Compute provided by our project sponsor Redmond AI, thank you! Follow RedmondAI on Twitter @RedmondAI.
16
+
17
+ ## Model Description
18
+
19
+ 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.
20
+
21
+ 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.
22
+
23
+ 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.
24
+
25
+ ## Example Outputs:
26
+ ![Example4](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example5.png "Example 4")
27
+ ![Example1](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/Example1.png "Example 1")
28
+ ![Example2](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example2.png "Example 2")
29
+ ![Example3](https://huggingface.co/NousResearch/Nous-Hermes-Llama2-13b/resolve/main/example3.png "Example 3")
30
+
31
+ ## Model Training
32
+
33
+ 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.
34
+
35
+ 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
36
+
37
+ ## Collaborators
38
+ The model fine-tuning and the datasets were a collaboration of efforts and resources between Teknium, Karan4D, Emozilla, Huemin Art, and Redmond AI.
39
+
40
+ Special mention goes to @winglian for assisting in some of the training issues.
41
+
42
+ Huge shoutout and acknowledgement is deserved for all the dataset creators who generously share their datasets openly.
43
+
44
+ Among the contributors of datasets:
45
+ - GPTeacher was made available by Teknium
46
+ - Wizard LM by nlpxucan
47
+ - Nous Research Instruct Dataset was provided by Karan4D and HueminArt.
48
+ - GPT4-LLM and Unnatural Instructions were provided by Microsoft
49
+ - Airoboros dataset by jondurbin
50
+ - Camel-AI's domain expert datasets are from Camel-AI
51
+ - CodeAlpaca dataset by Sahil 2801.
52
+
53
+ If anyone was left out, please open a thread in the community tab.
54
+
55
+ ## Prompt Format
56
+
57
+ The model follows the Alpaca prompt format:
58
+ ```
59
+ ### Instruction:
60
+ <prompt>
61
+
62
+ ### Response:
63
+ <leave a newline blank for model to respond>
64
+
65
+ ```
66
+
67
+ or
68
+
69
+ ```
70
+ ### Instruction:
71
+ <prompt>
72
+
73
+ ### Input:
74
+ <additional context>
75
+
76
+ ### Response:
77
+ <leave a newline blank for model to respond>
78
+
79
+ ```
80
+
81
+ ## Benchmark Results
82
+ AGI-Eval
83
+ ```
84
+ | Task |Version| Metric |Value | |Stderr|
85
+ |agieval_aqua_rat | 0|acc |0.2362|± |0.0267|
86
+ | | |acc_norm|0.2480|± |0.0272|
87
+ |agieval_logiqa_en | 0|acc |0.3425|± |0.0186|
88
+ | | |acc_norm|0.3472|± |0.0187|
89
+ |agieval_lsat_ar | 0|acc |0.2522|± |0.0287|
90
+ | | |acc_norm|0.2087|± |0.0269|
91
+ |agieval_lsat_lr | 0|acc |0.3510|± |0.0212|
92
+ | | |acc_norm|0.3627|± |0.0213|
93
+ |agieval_lsat_rc | 0|acc |0.4647|± |0.0305|
94
+ | | |acc_norm|0.4424|± |0.0303|
95
+ |agieval_sat_en | 0|acc |0.6602|± |0.0331|
96
+ | | |acc_norm|0.6165|± |0.0340|
97
+ |agieval_sat_en_without_passage| 0|acc |0.4320|± |0.0346|
98
+ | | |acc_norm|0.4272|± |0.0345|
99
+ |agieval_sat_math | 0|acc |0.2909|± |0.0307|
100
+ | | |acc_norm|0.2727|± |0.0301|
101
+ ```
102
+ GPT-4All Benchmark Set
103
+ ```
104
+ | Task |Version| Metric |Value | |Stderr|
105
+ |arc_challenge| 0|acc |0.5102|± |0.0146|
106
+ | | |acc_norm|0.5213|± |0.0146|
107
+ |arc_easy | 0|acc |0.7959|± |0.0083|
108
+ | | |acc_norm|0.7567|± |0.0088|
109
+ |boolq | 1|acc |0.8394|± |0.0064|
110
+ |hellaswag | 0|acc |0.6164|± |0.0049|
111
+ | | |acc_norm|0.8009|± |0.0040|
112
+ |openbookqa | 0|acc |0.3580|± |0.0215|
113
+ | | |acc_norm|0.4620|± |0.0223|
114
+ |piqa | 0|acc |0.7992|± |0.0093|
115
+ | | |acc_norm|0.8069|± |0.0092|
116
+ |winogrande | 0|acc |0.7127|± |0.0127|
117
+ ```
118
+ BigBench Reasoning Test
119
+ ```
120
+ | Task |Version| Metric |Value | |Stderr|
121
+
122
+ |bigbench_causal_judgement | 0|multiple_choice_grade|0.5526|± |0.0362|
123
+ |bigbench_date_understanding | 0|multiple_choice_grade|0.7344|± |0.0230|
124
+ |bigbench_disambiguation_qa | 0|multiple_choice_grade|0.2636|± |0.0275|
125
+ |bigbench_geometric_shapes | 0|multiple_choice_grade|0.0195|± |0.0073|
126
+ | | |exact_str_match |0.0000|± |0.0000|
127
+ |bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|0.2760|± |0.0200|
128
+ |bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|0.2100|± |0.0154|
129
+ |bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|0.4400|± |0.0287|
130
+ |bigbench_movie_recommendation | 0|multiple_choice_grade|0.2440|± |0.0192|
131
+ |bigbench_navigate | 0|multiple_choice_grade|0.4950|± |0.0158|
132
+ |bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|0.5570|± |0.0111|
133
+ |bigbench_ruin_names | 0|multiple_choice_grade|0.3728|± |0.0229|
134
+ |bigbench_salient_translation_error_detection | 0|multiple_choice_grade|0.1854|± |0.0123|
135
+ |bigbench_snarks | 0|multiple_choice_grade|0.6298|± |0.0360|
136
+ |bigbench_sports_understanding | 0|multiple_choice_grade|0.6156|± |0.0155|
137
+ |bigbench_temporal_sequences | 0|multiple_choice_grade|0.3140|± |0.0147|
138
+ |bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|0.2032|± |0.0114|
139
+ |bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|0.1406|± |0.0083|
140
+ |bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|0.4400|± |0.0287|
141
+ ```
142
+
143
+ These are the highest benchmarks Hermes has seen on every metric, achieving the following average scores:
144
+ - GPT4All benchmark average is now 70.0 - from 68.8 in Hermes-Llama1
145
+ - 0.3657 on BigBench, up from 0.328 on hermes-llama1
146
+ - 0.372 on AGIEval, up from 0.354 on Hermes-llama1
147
+
148
+ 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.
149
+
150
+ ## Resources for Applied Use Cases:
151
+ For an example of a back and forth chatbot using huggingface transformers and discord, check out: https://github.com/teknium1/alpaca-discord
152
+ For an example of a roleplaying discord chatbot, check out this: https://github.com/teknium1/alpaca-roleplay-discordbot
153
+
154
+ ## Future Plans
155
+ We plan to continue to iterate on both more high quality data, and new data filtering techniques to eliminate lower quality data going forward.
156
+
157
+ ## Model Usage
158
+ 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.
159
+
160
+ [<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)