TheBloke commited on
Commit
aa13b5d
1 Parent(s): 78ab354

Initial GGML model commit

Browse files
Files changed (1) hide show
  1. README.md +67 -40
README.md CHANGED
@@ -1,12 +1,6 @@
1
  ---
2
  inference: false
3
  license: other
4
- language:
5
- - en
6
- tags:
7
- - llama
8
- - self-instruct
9
- - distillation
10
  ---
11
 
12
  <!-- header start -->
@@ -38,41 +32,53 @@ GGML files are for CPU + GPU inference using [llama.cpp](https://github.com/gger
38
 
39
  * [4-bit GPTQ models for GPU inference](https://huggingface.co/TheBloke/Nous-Hermes-13B-GPTQ)
40
  * [4-bit, 5-bit, and 8-bit GGML models for CPU+GPU inference](https://huggingface.co/TheBloke/Nous-Hermes-13B-GGML)
41
- * [NousResearch's unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/NousResearch/Nous-Hermes-13b)
42
 
43
- ## Prompt Format
 
44
 
45
- The model follows the Alpaca prompt format:
46
- ```
47
- ### Instruction:
48
 
49
- ### Response:
50
- ```
51
 
52
- or
53
 
54
- ```
55
- ### Instruction:
56
 
57
- ### Input:
58
 
59
- ### Response:
60
- ```
61
 
62
- ## THE FILES IN MAIN BRANCH REQUIRES LATEST LLAMA.CPP (May 19th 2023 - commit 2d5db48)!
63
 
64
- llama.cpp recently made another breaking change to its quantisation methods - https://github.com/ggerganov/llama.cpp/pull/1508
 
 
 
 
 
 
65
 
66
- I have quantised the GGML files in this repo with the latest version. Therefore you will require llama.cpp compiled on May 19th or later (commit `2d5db48` or later) to use them.
 
67
 
68
  ## Provided files
69
  | Name | Quant method | Bits | Size | Max RAM required | Use case |
70
  | ---- | ---- | ---- | ---- | ---- | ----- |
71
- | nous-hermes-13b.ggmlv3.q4_0.bin | q4_0 | 4 | 7.32 GB | 9.82 GB | 4-bit. |
72
- | nous-hermes-13b.ggmlv3.q4_1.bin | q4_1 | 4 | 8.14 GB | 10.64 GB | 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
73
- | nous-hermes-13b.ggmlv3.q5_0.bin | q5_0 | 5 | 8.95 GB | 11.45 GB | 5-bit. Higher accuracy, higher resource usage and slower inference. |
74
- | nous-hermes-13b.ggmlv3.q5_1.bin | q5_1 | 5 | 9.76 GB | 12.26 GB | 5-bit. Even higher accuracy, resource usage and slower inference. |
75
- | nous-hermes-13b.ggmlv3.q8_0.bin | q8_0 | 8 | 13.83 GB | 16.33 GB | 8-bit. Almost indistinguishable from float16. Huge resource use and slow. Not recommended for normal use. |
 
 
 
 
 
 
 
 
 
76
 
77
 
78
  **Note**: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
@@ -114,35 +120,39 @@ Donaters will get priority support on any and all AI/LLM/model questions and req
114
  * Patreon: https://patreon.com/TheBlokeAI
115
  * Ko-Fi: https://ko-fi.com/TheBlokeAI
116
 
117
- **Patreon special mentions**: Aemon Algiz, Dmitriy Samsonov, Nathan LeClaire, Trenton Dambrowitz, Mano Prime, David Flickinger, vamX, Nikolai Manek, senxiiz, Khalefa Al-Ahmad, Illia Dulskyi, Jonathan Leane, Talal Aujan, V. Lukas, Joseph William Delisle, Pyrater, Oscar Rangel, Lone Striker, Luke Pendergrass, Eugene Pentland, Sebastain Graf, Johann-Peter Hartman.
 
 
118
 
119
  Thank you to all my generous patrons and donaters!
 
120
  <!-- footer end -->
121
 
122
  # Original model card: NousResearch's Nous-Hermes-13B
123
 
 
124
  # Model Card: Nous-Hermes-13b
125
 
126
  ## Model Description
127
 
128
  Nous-Hermes-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 Karan4D leading the fine tuning process and dataset curation, Redmond AI sponsoring the compute, and several other contributors. The result is an enhanced Llama 13b model that rivals GPT-3.5-turbo in performance across a variety of tasks.
129
 
130
- This model stands out for its long responses, low hallucination rate, and absence of OpenAI censorship mechanisms. The fine-tuning process was performed with a 2000 sequence length on an 8x a100 80GB DGX machine for over 50 hours.
131
 
132
  ## Model Training
133
 
134
- The model was trained almost entirely on synthetic GPT-4 outputs. This includes data from diverse sources such as GPTeacher, the general, roleplay v1&2, code instruct datasets, Nous Instruct & PDACTL (unpublished), CodeAlpaca, Evol_Instruct Uncensored, GPT4-LLM, and Unnatural Instructions.
135
 
136
  Additional data inputs came from Camel-AI's Biology/Physics/Chemistry and Math Datasets, Airoboros' GPT-4 Dataset, and more from CodeAlpaca. The total volume of data encompassed over 300,000 instructions.
137
 
138
  ## Collaborators
139
- The model fine-tuning and the datasets were a collaboration of efforts and resources between Teknium, Karan4D, Nous Research, Huemin Art, and Redmond AI.
140
-
141
- Huge shoutout and acknowledgement is deserved for all the dataset creators who generously share their datasets openly.
142
 
143
  Special mention goes to @winglian, @erhartford, and @main_horse for assisting in some of the training issues.
144
 
145
- Among the contributors of datasets, GPTeacher was made available by Teknium, Wizard LM by nlpxucan, and the Nous Research Instruct Dataset was provided by Karan4D and HueminArt.
146
  The GPT4-LLM and Unnatural Instructions were provided by Microsoft, Airoboros dataset by jondurbin, Camel-AI datasets are from Camel-AI, and CodeAlpaca dataset by Sahil 2801.
147
  If anyone was left out, please open a thread in the community tab.
148
 
@@ -155,7 +165,7 @@ The model follows the Alpaca prompt format:
155
  ### Response:
156
  ```
157
 
158
- or
159
 
160
  ```
161
  ### Instruction:
@@ -163,19 +173,36 @@ or
163
  ### Input:
164
 
165
  ### Response:
166
- ```
167
 
168
  ## Resources for Applied Use Cases:
169
- For an example of a back and forth chatbot using huggingface transformers and discord, check out: https://github.com/teknium1/alpaca-discord
170
- For an example of a roleplaying discord bot, check out this: https://github.com/teknium1/alpaca-roleplay-discordbot
171
 
172
  ## Future Plans
173
  The model is currently being uploaded in FP16 format, and there are plans to convert the model to GGML and GPTQ 4bit quantizations. The team is also working on a full benchmark, similar to what was done for GPT4-x-Vicuna. We will try to get in discussions to get the model included in the GPT4All.
174
 
175
  ## Benchmark Results
176
- Benchmark results are coming soon.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
177
 
178
  ## Model Usage
179
  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.
180
-
181
  Compute provided by our project sponsor Redmond AI, thank you!!
 
1
  ---
2
  inference: false
3
  license: other
 
 
 
 
 
 
4
  ---
5
 
6
  <!-- header start -->
 
32
 
33
  * [4-bit GPTQ models for GPU inference](https://huggingface.co/TheBloke/Nous-Hermes-13B-GPTQ)
34
  * [4-bit, 5-bit, and 8-bit GGML models for CPU+GPU inference](https://huggingface.co/TheBloke/Nous-Hermes-13B-GGML)
35
+ * [Unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/NousResearch/Nous-Hermes-13b)
36
 
37
+ <!-- compatibility_ggml start -->
38
+ ## Compatibility
39
 
40
+ ### Original llama.cpp quant methods: `q4_0, q4_1, q5_0, q5_1, q8_0`
 
 
41
 
42
+ I have quantized these 'original' quantisation methods using an older version of llama.cpp so that they remain compatible with llama.cpp as of May 19th, commit `2d5db48`.
 
43
 
44
+ They should be compatible with all current UIs and libraries that use llama.cpp, such as those listed at the top of this README.
45
 
46
+ ### New k-quant methods: `q2_K, q3_K_S, q3_K_M, q3_K_L, q4_K_S, q4_K_M, q5_K_S, q6_K`
 
47
 
48
+ These new quantisation methods are only compatible with llama.cpp as of June 6th, commit `2d43387`.
49
 
50
+ They will NOT be compatible with koboldcpp, text-generation-ui, and other UIs and libraries yet. Support is expected to come over the next few days.
 
51
 
52
+ ## Explanation of the new k-quant methods
53
 
54
+ The new methods available are:
55
+ * GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
56
+ * GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
57
+ * GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
58
+ * GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
59
+ * GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
60
+ * GGML_TYPE_Q8_K - "type-0" 8-bit quantization. Only used for quantizing intermediate results. The difference to the existing Q8_0 is that the block size is 256. All 2-6 bit dot products are implemented for this quantization type.
61
 
62
+ Refer to the Provided Files table below to see what files use which methods, and how.
63
+ <!-- compatibility_ggml end -->
64
 
65
  ## Provided files
66
  | Name | Quant method | Bits | Size | Max RAM required | Use case |
67
  | ---- | ---- | ---- | ---- | ---- | ----- |
68
+ | nous-hermes-13b.ggmlv3.q2_K.bin | q2_K | 2 | 5.43 GB | 7.93 GB | New k-quant method. Uses GGML_TYPE_Q4_K for the attention.vw and feed_forward.w2 tensors, GGML_TYPE_Q2_K for the other tensors. |
69
+ | nous-hermes-13b.ggmlv3.q3_K_L.bin | q3_K_L | 3 | 6.87 GB | 9.37 GB | New k-quant method. Uses GGML_TYPE_Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K |
70
+ | nous-hermes-13b.ggmlv3.q3_K_M.bin | q3_K_M | 3 | 6.25 GB | 8.75 GB | New k-quant method. Uses GGML_TYPE_Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K |
71
+ | nous-hermes-13b.ggmlv3.q3_K_S.bin | q3_K_S | 3 | 5.59 GB | 8.09 GB | New k-quant method. Uses GGML_TYPE_Q3_K for all tensors |
72
+ | nous-hermes-13b.ggmlv3.q4_0.bin | q4_0 | 4 | 7.32 GB | 9.82 GB | Original llama.cpp quant method, 4-bit. |
73
+ | nous-hermes-13b.ggmlv3.q4_1.bin | q4_1 | 4 | 8.14 GB | 10.64 GB | Original llama.cpp quant method, 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
74
+ | nous-hermes-13b.ggmlv3.q4_K_M.bin | q4_K_M | 4 | 7.82 GB | 10.32 GB | New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q4_K |
75
+ | nous-hermes-13b.ggmlv3.q4_K_S.bin | q4_K_S | 4 | 7.32 GB | 9.82 GB | New k-quant method. Uses GGML_TYPE_Q4_K for all tensors |
76
+ | nous-hermes-13b.ggmlv3.q5_0.bin | q5_0 | 5 | 8.95 GB | 11.45 GB | Original llama.cpp quant method, 5-bit. Higher accuracy, higher resource usage and slower inference. |
77
+ | nous-hermes-13b.ggmlv3.q5_1.bin | q5_1 | 5 | 9.76 GB | 12.26 GB | Original llama.cpp quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
78
+ | nous-hermes-13b.ggmlv3.q5_K_M.bin | q5_K_M | 5 | 9.21 GB | 11.71 GB | New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q5_K |
79
+ | nous-hermes-13b.ggmlv3.q5_K_S.bin | q5_K_S | 5 | 8.95 GB | 11.45 GB | New k-quant method. Uses GGML_TYPE_Q5_K for all tensors |
80
+ | nous-hermes-13b.ggmlv3.q6_K.bin | q6_K | 6 | 10.68 GB | 13.18 GB | New k-quant method. Uses GGML_TYPE_Q8_K - 6-bit quantization - for all tensors |
81
+ | nous-hermes-13b.ggmlv3.q8_0.bin | q8_0 | 8 | 13.83 GB | 16.33 GB | Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
82
 
83
 
84
  **Note**: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
 
120
  * Patreon: https://patreon.com/TheBlokeAI
121
  * Ko-Fi: https://ko-fi.com/TheBlokeAI
122
 
123
+ **Special thanks to**: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.
124
+
125
+ **Patreon special mentions**: Ajan Kanaga, Kalila, Derek Yates, Sean Connelly, Luke, Nathan LeClaire, Trenton Dambrowitz, Mano Prime, David Flickinger, vamX, Nikolai Manek, senxiiz, Khalefa Al-Ahmad, Illia Dulskyi, trip7s trip, Jonathan Leane, Talal Aujan, Artur Olbinski, Cory Kujawski, Joseph William Delisle, Pyrater, Oscar Rangel, Lone Striker, Luke Pendergrass, Eugene Pentland, Johann-Peter Hartmann.
126
 
127
  Thank you to all my generous patrons and donaters!
128
+
129
  <!-- footer end -->
130
 
131
  # Original model card: NousResearch's Nous-Hermes-13B
132
 
133
+
134
  # Model Card: Nous-Hermes-13b
135
 
136
  ## Model Description
137
 
138
  Nous-Hermes-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 Karan4D leading the fine tuning process and dataset curation, Redmond AI sponsoring the compute, and several other contributors. The result is an enhanced Llama 13b model that rivals GPT-3.5-turbo in performance across a variety of tasks.
139
 
140
+ This model stands out for its long responses, low hallucination rate, and absence of OpenAI censorship mechanisms. The fine-tuning process was performed with a 2000 sequence length on an 8x a100 80GB DGX machine for over 50 hours.
141
 
142
  ## Model Training
143
 
144
+ The model was trained almost entirely on synthetic GPT-4 outputs. This includes data from diverse sources such as GPTeacher, the general, roleplay v1&2, code instruct datasets, Nous Instruct & PDACTL (unpublished), CodeAlpaca, Evol_Instruct Uncensored, GPT4-LLM, and Unnatural Instructions.
145
 
146
  Additional data inputs came from Camel-AI's Biology/Physics/Chemistry and Math Datasets, Airoboros' GPT-4 Dataset, and more from CodeAlpaca. The total volume of data encompassed over 300,000 instructions.
147
 
148
  ## Collaborators
149
+ The model fine-tuning and the datasets were a collaboration of efforts and resources between Teknium, Karan4D, Nous Research, Huemin Art, and Redmond AI.
150
+
151
+ Huge shoutout and acknowledgement is deserved for all the dataset creators who generously share their datasets openly.
152
 
153
  Special mention goes to @winglian, @erhartford, and @main_horse for assisting in some of the training issues.
154
 
155
+ Among the contributors of datasets, GPTeacher was made available by Teknium, Wizard LM by nlpxucan, and the Nous Research Instruct Dataset was provided by Karan4D and HueminArt.
156
  The GPT4-LLM and Unnatural Instructions were provided by Microsoft, Airoboros dataset by jondurbin, Camel-AI datasets are from Camel-AI, and CodeAlpaca dataset by Sahil 2801.
157
  If anyone was left out, please open a thread in the community tab.
158
 
 
165
  ### Response:
166
  ```
167
 
168
+ or
169
 
170
  ```
171
  ### Instruction:
 
173
  ### Input:
174
 
175
  ### Response:
176
+ ```
177
 
178
  ## Resources for Applied Use Cases:
179
+ For an example of a back and forth chatbot using huggingface transformers and discord, check out: https://github.com/teknium1/alpaca-discord
180
+ For an example of a roleplaying discord bot, check out this: https://github.com/teknium1/alpaca-roleplay-discordbot
181
 
182
  ## Future Plans
183
  The model is currently being uploaded in FP16 format, and there are plans to convert the model to GGML and GPTQ 4bit quantizations. The team is also working on a full benchmark, similar to what was done for GPT4-x-Vicuna. We will try to get in discussions to get the model included in the GPT4All.
184
 
185
  ## Benchmark Results
186
+ ```
187
+ | Task |Version| Metric |Value | |Stderr|
188
+ |-------------|------:|--------|-----:|---|-----:|
189
+ |arc_challenge| 0|acc |0.4915|± |0.0146|
190
+ | | |acc_norm|0.5085|± |0.0146|
191
+ |arc_easy | 0|acc |0.7769|± |0.0085|
192
+ | | |acc_norm|0.7424|± |0.0090|
193
+ |boolq | 1|acc |0.7948|± |0.0071|
194
+ |hellaswag | 0|acc |0.6143|± |0.0049|
195
+ | | |acc_norm|0.8000|± |0.0040|
196
+ |openbookqa | 0|acc |0.3560|± |0.0214|
197
+ | | |acc_norm|0.4640|± |0.0223|
198
+ |piqa | 0|acc |0.7965|± |0.0094|
199
+ | | |acc_norm|0.7889|± |0.0095|
200
+ |winogrande | 0|acc |0.7190|± |0.0126|
201
+ ```
202
+
203
+ 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.
204
 
205
  ## Model Usage
206
  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.
207
+
208
  Compute provided by our project sponsor Redmond AI, thank you!!