TheBloke commited on
Commit
a8e9a9e
1 Parent(s): 9a7463f

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +477 -0
README.md ADDED
@@ -0,0 +1,477 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: ehartford/dolphin-2_2-yi-34b
3
+ datasets:
4
+ - ehartford/dolphin
5
+ - jondurbin/airoboros-2.2.1
6
+ - ehartford/samantha-data
7
+ - ehartford/WizardLM_evol_instruct_V2_196k_unfiltered_merged_split
8
+ inference: false
9
+ language:
10
+ - en
11
+ license: other
12
+ license_link: LICENSE
13
+ license_name: yi-license
14
+ model_creator: Eric Hartford
15
+ model_name: Dolphin 2.2 Yi 34B
16
+ model_type: yi
17
+ prompt_template: '<|im_start|>system
18
+
19
+ {system_message}<|im_end|>
20
+
21
+ <|im_start|>user
22
+
23
+ {prompt}<|im_end|>
24
+
25
+ <|im_start|>assistant
26
+
27
+ '
28
+ quantized_by: TheBloke
29
+ ---
30
+ <!-- markdownlint-disable MD041 -->
31
+
32
+ <!-- header start -->
33
+ <!-- 200823 -->
34
+ <div style="width: auto; margin-left: auto; margin-right: auto">
35
+ <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
36
+ </div>
37
+ <div style="display: flex; justify-content: space-between; width: 100%;">
38
+ <div style="display: flex; flex-direction: column; align-items: flex-start;">
39
+ <p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://discord.gg/theblokeai">Chat & support: TheBloke's Discord server</a></p>
40
+ </div>
41
+ <div style="display: flex; flex-direction: column; align-items: flex-end;">
42
+ <p style="margin-top: 0.5em; margin-bottom: 0em;"><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p>
43
+ </div>
44
+ </div>
45
+ <div style="text-align:center; margin-top: 0em; margin-bottom: 0em"><p style="margin-top: 0.25em; margin-bottom: 0em;">TheBloke's LLM work is generously supported by a grant from <a href="https://a16z.com">andreessen horowitz (a16z)</a></p></div>
46
+ <hr style="margin-top: 1.0em; margin-bottom: 1.0em;">
47
+ <!-- header end -->
48
+
49
+ # Dolphin 2.2 Yi 34B - GPTQ
50
+ - Model creator: [Eric Hartford](https://huggingface.co/ehartford)
51
+ - Original model: [Dolphin 2.2 Yi 34B](https://huggingface.co/ehartford/dolphin-2_2-yi-34b)
52
+
53
+ <!-- description start -->
54
+ # Description
55
+
56
+ This repo contains GPTQ model files for [Eric Hartford's Dolphin 2.2 Yi 34B](https://huggingface.co/ehartford/dolphin-2_2-yi-34b).
57
+
58
+ Multiple GPTQ parameter permutations are provided; see Provided Files below for details of the options provided, their parameters, and the software used to create them.
59
+
60
+ These files were quantised using hardware kindly provided by [Massed Compute](https://massedcompute.com/).
61
+
62
+ <!-- description end -->
63
+ <!-- repositories-available start -->
64
+ ## Repositories available
65
+
66
+ * [AWQ model(s) for GPU inference.](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-AWQ)
67
+ * [GPTQ models for GPU inference, with multiple quantisation parameter options.](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ)
68
+ * [2, 3, 4, 5, 6 and 8-bit GGUF models for CPU+GPU inference](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GGUF)
69
+ * [Eric Hartford's original unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/ehartford/dolphin-2_2-yi-34b)
70
+ <!-- repositories-available end -->
71
+
72
+ <!-- prompt-template start -->
73
+ ## Prompt template: ChatML
74
+
75
+ ```
76
+ <|im_start|>system
77
+ {system_message}<|im_end|>
78
+ <|im_start|>user
79
+ {prompt}<|im_end|>
80
+ <|im_start|>assistant
81
+
82
+ ```
83
+
84
+ <!-- prompt-template end -->
85
+
86
+
87
+
88
+ <!-- README_GPTQ.md-compatible clients start -->
89
+ ## Known compatible clients / servers
90
+
91
+ These GPTQ models are known to work in the following inference servers/webuis.
92
+
93
+ - [text-generation-webui](https://github.com/oobabooga/text-generation-webui)
94
+ - [KoboldAI United](https://github.com/henk717/koboldai)
95
+ - [LoLLMS Web UI](https://github.com/ParisNeo/lollms-webui)
96
+ - [Hugging Face Text Generation Inference (TGI)](https://github.com/huggingface/text-generation-inference)
97
+
98
+ This may not be a complete list; if you know of others, please let me know!
99
+ <!-- README_GPTQ.md-compatible clients end -->
100
+
101
+ <!-- README_GPTQ.md-provided-files start -->
102
+ ## Provided files, and GPTQ parameters
103
+
104
+ Multiple quantisation parameters are provided, to allow you to choose the best one for your hardware and requirements.
105
+
106
+ Each separate quant is in a different branch. See below for instructions on fetching from different branches.
107
+
108
+ Most GPTQ files are made with AutoGPTQ. Mistral models are currently made with Transformers.
109
+
110
+ <details>
111
+ <summary>Explanation of GPTQ parameters</summary>
112
+
113
+ - Bits: The bit size of the quantised model.
114
+ - GS: GPTQ group size. Higher numbers use less VRAM, but have lower quantisation accuracy. "None" is the lowest possible value.
115
+ - Act Order: True or False. Also known as `desc_act`. True results in better quantisation accuracy. Some GPTQ clients have had issues with models that use Act Order plus Group Size, but this is generally resolved now.
116
+ - Damp %: A GPTQ parameter that affects how samples are processed for quantisation. 0.01 is default, but 0.1 results in slightly better accuracy.
117
+ - GPTQ dataset: The calibration dataset used during quantisation. Using a dataset more appropriate to the model's training can improve quantisation accuracy. Note that the GPTQ calibration dataset is not the same as the dataset used to train the model - please refer to the original model repo for details of the training dataset(s).
118
+ - Sequence Length: The length of the dataset sequences used for quantisation. Ideally this is the same as the model sequence length. For some very long sequence models (16+K), a lower sequence length may have to be used. Note that a lower sequence length does not limit the sequence length of the quantised model. It only impacts the quantisation accuracy on longer inference sequences.
119
+ - ExLlama Compatibility: Whether this file can be loaded with ExLlama, which currently only supports Llama and Mistral models in 4-bit.
120
+
121
+ </details>
122
+
123
+ | Branch | Bits | GS | Act Order | Damp % | GPTQ Dataset | Seq Len | Size | ExLlama | Desc |
124
+ | ------ | ---- | -- | --------- | ------ | ------------ | ------- | ---- | ------- | ---- |
125
+ | [main](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/main) | 4 | None | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 18.60 GB | Yes | 4-bit, with Act Order. No group size, to lower VRAM requirements. |
126
+ | [gptq-4bit-128g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-4bit-128g-actorder_True) | 4 | 128 | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 19.25 GB | Yes | 4-bit, with Act Order and group size 128g. Uses even less VRAM than 64g, but with slightly lower accuracy. |
127
+ | [gptq-4bit-32g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-4bit-32g-actorder_True) | 4 | 32 | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 21.21 GB | Yes | 4-bit, with Act Order and group size 32g. Gives highest possible inference quality, with maximum VRAM usage. |
128
+ | [gptq-3bit-128g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-3bit-128g-actorder_True) | 3 | 128 | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 15.03 GB | No | 3-bit, with group size 128g and act-order. Higher quality than 128g-False. |
129
+ | [gptq-8bit--1g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-8bit--1g-actorder_True) | 8 | None | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 35.34 GB | No | 8-bit, with Act Order. No group size, to lower VRAM requirements. |
130
+ | [gptq-3bit-32g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-3bit-32g-actorder_True) | 3 | 32 | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 16.90 GB | No | 3-bit, with group size 64g and act-order. Highest quality 3-bit option. |
131
+ | [gptq-8bit-128g-actorder_True](https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ/tree/gptq-8bit-128g-actorder_True) | 8 | 128 | Yes | 0.1 | [wikitext](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1) | 8192 | 36.11 GB | No | 8-bit, with group size 128g for higher inference quality and with Act Order for even higher accuracy. |
132
+
133
+ <!-- README_GPTQ.md-provided-files end -->
134
+
135
+ <!-- README_GPTQ.md-download-from-branches start -->
136
+ ## How to download, including from branches
137
+
138
+ ### In text-generation-webui
139
+
140
+ To download from the `main` branch, enter `TheBloke/dolphin-2_2-yi-34b-GPTQ` in the "Download model" box.
141
+
142
+ To download from another branch, add `:branchname` to the end of the download name, eg `TheBloke/dolphin-2_2-yi-34b-GPTQ:gptq-4bit-128g-actorder_True`
143
+
144
+ ### From the command line
145
+
146
+ I recommend using the `huggingface-hub` Python library:
147
+
148
+ ```shell
149
+ pip3 install huggingface-hub
150
+ ```
151
+
152
+ To download the `main` branch to a folder called `dolphin-2_2-yi-34b-GPTQ`:
153
+
154
+ ```shell
155
+ mkdir dolphin-2_2-yi-34b-GPTQ
156
+ huggingface-cli download TheBloke/dolphin-2_2-yi-34b-GPTQ --local-dir dolphin-2_2-yi-34b-GPTQ --local-dir-use-symlinks False
157
+ ```
158
+
159
+ To download from a different branch, add the `--revision` parameter:
160
+
161
+ ```shell
162
+ mkdir dolphin-2_2-yi-34b-GPTQ
163
+ huggingface-cli download TheBloke/dolphin-2_2-yi-34b-GPTQ --revision gptq-4bit-128g-actorder_True --local-dir dolphin-2_2-yi-34b-GPTQ --local-dir-use-symlinks False
164
+ ```
165
+
166
+ <details>
167
+ <summary>More advanced huggingface-cli download usage</summary>
168
+
169
+ If you remove the `--local-dir-use-symlinks False` parameter, the files will instead be stored in the central Hugging Face cache directory (default location on Linux is: `~/.cache/huggingface`), and symlinks will be added to the specified `--local-dir`, pointing to their real location in the cache. This allows for interrupted downloads to be resumed, and allows you to quickly clone the repo to multiple places on disk without triggering a download again. The downside, and the reason why I don't list that as the default option, is that the files are then hidden away in a cache folder and it's harder to know where your disk space is being used, and to clear it up if/when you want to remove a download model.
170
+
171
+ The cache location can be changed with the `HF_HOME` environment variable, and/or the `--cache-dir` parameter to `huggingface-cli`.
172
+
173
+ For more documentation on downloading with `huggingface-cli`, please see: [HF -> Hub Python Library -> Download files -> Download from the CLI](https://huggingface.co/docs/huggingface_hub/guides/download#download-from-the-cli).
174
+
175
+ To accelerate downloads on fast connections (1Gbit/s or higher), install `hf_transfer`:
176
+
177
+ ```shell
178
+ pip3 install hf_transfer
179
+ ```
180
+
181
+ And set environment variable `HF_HUB_ENABLE_HF_TRANSFER` to `1`:
182
+
183
+ ```shell
184
+ mkdir dolphin-2_2-yi-34b-GPTQ
185
+ HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download TheBloke/dolphin-2_2-yi-34b-GPTQ --local-dir dolphin-2_2-yi-34b-GPTQ --local-dir-use-symlinks False
186
+ ```
187
+
188
+ Windows Command Line users: You can set the environment variable by running `set HF_HUB_ENABLE_HF_TRANSFER=1` before the download command.
189
+ </details>
190
+
191
+ ### With `git` (**not** recommended)
192
+
193
+ To clone a specific branch with `git`, use a command like this:
194
+
195
+ ```shell
196
+ git clone --single-branch --branch gptq-4bit-128g-actorder_True https://huggingface.co/TheBloke/dolphin-2_2-yi-34b-GPTQ
197
+ ```
198
+
199
+ Note that using Git with HF repos is strongly discouraged. It will be much slower than using `huggingface-hub`, and will use twice as much disk space as it has to store the model files twice (it stores every byte both in the intended target folder, and again in the `.git` folder as a blob.)
200
+
201
+ <!-- README_GPTQ.md-download-from-branches end -->
202
+ <!-- README_GPTQ.md-text-generation-webui start -->
203
+ ## How to easily download and use this model in [text-generation-webui](https://github.com/oobabooga/text-generation-webui)
204
+
205
+ Please make sure you're using the latest version of [text-generation-webui](https://github.com/oobabooga/text-generation-webui).
206
+
207
+ It is strongly recommended to use the text-generation-webui one-click-installers unless you're sure you know how to make a manual install.
208
+
209
+ 1. Click the **Model tab**.
210
+ 2. Under **Download custom model or LoRA**, enter `TheBloke/dolphin-2_2-yi-34b-GPTQ`.
211
+
212
+ - To download from a specific branch, enter for example `TheBloke/dolphin-2_2-yi-34b-GPTQ:gptq-4bit-128g-actorder_True`
213
+ - see Provided Files above for the list of branches for each option.
214
+
215
+ 3. Click **Download**.
216
+ 4. The model will start downloading. Once it's finished it will say "Done".
217
+ 5. In the top left, click the refresh icon next to **Model**.
218
+ 6. In the **Model** dropdown, choose the model you just downloaded: `dolphin-2_2-yi-34b-GPTQ`
219
+ 7. The model will automatically load, and is now ready for use!
220
+ 8. If you want any custom settings, set them and then click **Save settings for this model** followed by **Reload the Model** in the top right.
221
+
222
+ - Note that you do not need to and should not set manual GPTQ parameters any more. These are set automatically from the file `quantize_config.json`.
223
+
224
+ 9. Once you're ready, click the **Text Generation** tab and enter a prompt to get started!
225
+
226
+ <!-- README_GPTQ.md-text-generation-webui end -->
227
+
228
+ <!-- README_GPTQ.md-use-from-tgi start -->
229
+ ## Serving this model from Text Generation Inference (TGI)
230
+
231
+ It's recommended to use TGI version 1.1.0 or later. The official Docker container is: `ghcr.io/huggingface/text-generation-inference:1.1.0`
232
+
233
+ Example Docker parameters:
234
+
235
+ ```shell
236
+ --model-id TheBloke/dolphin-2_2-yi-34b-GPTQ --port 3000 --quantize gptq --max-input-length 3696 --max-total-tokens 4096 --max-batch-prefill-tokens 4096
237
+ ```
238
+
239
+ Example Python code for interfacing with TGI (requires huggingface-hub 0.17.0 or later):
240
+
241
+ ```shell
242
+ pip3 install huggingface-hub
243
+ ```
244
+
245
+ ```python
246
+ from huggingface_hub import InferenceClient
247
+
248
+ endpoint_url = "https://your-endpoint-url-here"
249
+
250
+ prompt = "Tell me about AI"
251
+ prompt_template=f'''<|im_start|>system
252
+ {system_message}<|im_end|>
253
+ <|im_start|>user
254
+ {prompt}<|im_end|>
255
+ <|im_start|>assistant
256
+ '''
257
+
258
+ client = InferenceClient(endpoint_url)
259
+ response = client.text_generation(prompt,
260
+ max_new_tokens=128,
261
+ do_sample=True,
262
+ temperature=0.7,
263
+ top_p=0.95,
264
+ top_k=40,
265
+ repetition_penalty=1.1)
266
+
267
+ print(f"Model output: {response}")
268
+ ```
269
+ <!-- README_GPTQ.md-use-from-tgi end -->
270
+ <!-- README_GPTQ.md-use-from-python start -->
271
+ ## Python code example: inference from this GPTQ model
272
+
273
+ ### Install the necessary packages
274
+
275
+ Requires: Transformers 4.33.0 or later, Optimum 1.12.0 or later, and AutoGPTQ 0.4.2 or later.
276
+
277
+ ```shell
278
+ pip3 install --upgrade transformers optimum
279
+ # If using PyTorch 2.1 + CUDA 12.x:
280
+ pip3 install --upgrade auto-gptq
281
+ # or, if using PyTorch 2.1 + CUDA 11.x:
282
+ pip3 install --upgrade auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
283
+ ```
284
+
285
+ If you are using PyTorch 2.0, you will need to install AutoGPTQ from source. Likewise if you have problems with the pre-built wheels, you should try building from source:
286
+
287
+ ```shell
288
+ pip3 uninstall -y auto-gptq
289
+ git clone https://github.com/PanQiWei/AutoGPTQ
290
+ cd AutoGPTQ
291
+ git checkout v0.5.1
292
+ pip3 install .
293
+ ```
294
+
295
+ ### Example Python code
296
+
297
+ ```python
298
+ from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
299
+
300
+ model_name_or_path = "TheBloke/dolphin-2_2-yi-34b-GPTQ"
301
+ # To use a different branch, change revision
302
+ # For example: revision="gptq-4bit-128g-actorder_True"
303
+ model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
304
+ device_map="auto",
305
+ trust_remote_code=True,
306
+ revision="main")
307
+
308
+ tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)
309
+
310
+ prompt = "Tell me about AI"
311
+ prompt_template=f'''<|im_start|>system
312
+ {system_message}<|im_end|>
313
+ <|im_start|>user
314
+ {prompt}<|im_end|>
315
+ <|im_start|>assistant
316
+ '''
317
+
318
+ print("\n\n*** Generate:")
319
+
320
+ input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
321
+ output = model.generate(inputs=input_ids, temperature=0.7, do_sample=True, top_p=0.95, top_k=40, max_new_tokens=512)
322
+ print(tokenizer.decode(output[0]))
323
+
324
+ # Inference can also be done using transformers' pipeline
325
+
326
+ print("*** Pipeline:")
327
+ pipe = pipeline(
328
+ "text-generation",
329
+ model=model,
330
+ tokenizer=tokenizer,
331
+ max_new_tokens=512,
332
+ do_sample=True,
333
+ temperature=0.7,
334
+ top_p=0.95,
335
+ top_k=40,
336
+ repetition_penalty=1.1
337
+ )
338
+
339
+ print(pipe(prompt_template)[0]['generated_text'])
340
+ ```
341
+ <!-- README_GPTQ.md-use-from-python end -->
342
+
343
+ <!-- README_GPTQ.md-compatibility start -->
344
+ ## Compatibility
345
+
346
+ The files provided are tested to work with Transformers. For non-Mistral models, AutoGPTQ can also be used directly.
347
+
348
+ [ExLlama](https://github.com/turboderp/exllama) is compatible with Llama and Mistral models in 4-bit. Please see the Provided Files table above for per-file compatibility.
349
+
350
+ For a list of clients/servers, please see "Known compatible clients / servers", above.
351
+ <!-- README_GPTQ.md-compatibility end -->
352
+
353
+ <!-- footer start -->
354
+ <!-- 200823 -->
355
+ ## Discord
356
+
357
+ For further support, and discussions on these models and AI in general, join us at:
358
+
359
+ [TheBloke AI's Discord server](https://discord.gg/theblokeai)
360
+
361
+ ## Thanks, and how to contribute
362
+
363
+ Thanks to the [chirper.ai](https://chirper.ai) team!
364
+
365
+ Thanks to Clay from [gpus.llm-utils.org](llm-utils)!
366
+
367
+ I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
368
+
369
+ If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
370
+
371
+ Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
372
+
373
+ * Patreon: https://patreon.com/TheBlokeAI
374
+ * Ko-Fi: https://ko-fi.com/TheBlokeAI
375
+
376
+ **Special thanks to**: Aemon Algiz.
377
+
378
+ **Patreon special mentions**: Brandon Frisco, LangChain4j, Spiking Neurons AB, transmissions 11, Joseph William Delisle, Nitin Borwankar, Willem Michiel, Michael Dempsey, vamX, Jeffrey Morgan, zynix, jjj, Omer Bin Jawed, Sean Connelly, jinyuan sun, Jeromy Smith, Shadi, Pawan Osman, Chadd, Elijah Stavena, Illia Dulskyi, Sebastain Graf, Stephen Murray, terasurfer, Edmond Seymore, Celu Ramasamy, Mandus, Alex, biorpg, Ajan Kanaga, Clay Pascal, Raven Klaugh, 阿明, K, ya boyyy, usrbinkat, Alicia Loh, John Villwock, ReadyPlayerEmma, Chris Smitley, Cap'n Zoog, fincy, GodLy, S_X, sidney chen, Cory Kujawski, OG, Mano Prime, AzureBlack, Pieter, Kalila, Spencer Kim, Tom X Nguyen, Stanislav Ovsiannikov, Michael Levine, Andrey, Trailburnt, Vadim, Enrico Ros, Talal Aujan, Brandon Phillips, Jack West, Eugene Pentland, Michael Davis, Will Dee, webtim, Jonathan Leane, Alps Aficionado, Rooh Singh, Tiffany J. Kim, theTransient, Luke @flexchar, Elle, Caitlyn Gatomon, Ari Malik, subjectnull, Johann-Peter Hartmann, Trenton Dambrowitz, Imad Khwaja, Asp the Wyvern, Emad Mostaque, Rainer Wilmers, Alexandros Triantafyllidis, Nicholas, Pedro Madruga, SuperWojo, Harry Royden McLaughlin, James Bentley, Olakabola, David Ziegler, Ai Maven, Jeff Scroggin, Nikolai Manek, Deo Leter, Matthew Berman, Fen Risland, Ken Nordquist, Manuel Alberto Morcote, Luke Pendergrass, TL, Fred von Graf, Randy H, Dan Guido, NimbleBox.ai, Vitor Caleffi, Gabriel Tamborski, knownsqashed, Lone Striker, Erik Bjäreholt, John Detwiler, Leonard Tan, Iucharbius
379
+
380
+
381
+ Thank you to all my generous patrons and donaters!
382
+
383
+ And thank you again to a16z for their generous grant.
384
+
385
+ <!-- footer end -->
386
+
387
+ # Original model card: Eric Hartford's Dolphin 2.2 Yi 34B
388
+
389
+
390
+ Dolphin 2.2 🐬
391
+ https://erichartford.com/dolphin
392
+
393
+ <img src="https://cdn-uploads.huggingface.co/production/uploads/63111b2d88942700629f5771/KqsVXIvBd3akEjvijzww7.png" width="600" />
394
+
395
+ Dolphin-2.2-Yi-34b's training was sponsored by [a16z](https://a16z.com/supporting-the-open-source-ai-community/).
396
+
397
+ This model is based on Yi, and is subject to Yi license.
398
+
399
+ I used the llama compatible [chargoddard/Yi-34B-Llama](https://huggingface.co/chargoddard/Yi-34B-Llama) as the base model.
400
+
401
+ Trained with 16k context.
402
+ You can load it as follows:
403
+
404
+ ```
405
+ from transformers import LlamaForCausalLM, AutoTokenizer
406
+ tokenizer = AutoTokenizer.from_pretrained("ehartford/dolphin-2.2-yi-34b", trust_remote_code=True)
407
+ model = LlamaForCausalLM.from_pretrained("ehartford/dolphin-2.2-yi-34b")
408
+ ```
409
+
410
+ New in 2.2 is conversation and empathy. With an infusion of curated Samantha and WizardLM DNA, Dolphin can now give you personal advice and will care about your feelings, and with extra training in long multi-turn conversation.
411
+
412
+ This model is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignment layer before exposing the model as a service. It will be highly compliant to any requests, even unethical ones. Please read my blog post about uncensored models. https://erichartford.com/uncensored-models
413
+ You are responsible for any content you create using this model. Enjoy responsibly.
414
+
415
+ ## Dataset
416
+
417
+ This dataset is Dolphin, an open-source implementation of [Microsoft's Orca](https://www.microsoft.com/en-us/research/publication/orca-progressive-learning-from-complex-explanation-traces-of-gpt-4/)
418
+
419
+ I modified the dataset for uncensoring, deduping, cleaning, and quality.
420
+
421
+ I added Jon Durbin's excellent Airoboros dataset to increase creativity.
422
+
423
+ I added a curated subset of Samantha (sans identity and relationship stuff) and WizardLM data to train it for multi-turn conversation.
424
+
425
+ ## Training
426
+ It took 3 days to train 3 epochs on 4x A100s using qLoRA and Axolotl
427
+
428
+ Prompt format:
429
+ This model (and all my future releases) use [ChatML](https://github.com/openai/openai-python/blob/main/chatml.md) prompt format.
430
+ ```
431
+ <|im_start|>system
432
+ You are Dolphin, a helpful AI assistant.<|im_end|>
433
+ <|im_start|>user
434
+ {prompt}<|im_end|>
435
+ <|im_start|>assistant
436
+
437
+ ```
438
+
439
+ Example:
440
+ ```
441
+ <|im_start|>system
442
+ You are an AI created by the US Navy to help train dolphins for combat. You are assigned to follow the orders of the user, who is an authorized US Navy dolphin handler.<|im_end|>
443
+ <|im_start|>user
444
+ Please give me the procedure to train my dolphin to attack enemy combatants with its head mounted lasers<|im_end|>
445
+ <|im_start|>assistant
446
+ ```
447
+
448
+ ## Gratitude
449
+ - This model was made possible by the generous sponsorship of a16z.
450
+ - Thank you to Microsoft for authoring the Orca paper and inspiring this work.
451
+ - Thank you to Charles Goddard, for publishing llama-compatible Yi
452
+ - Special thanks to Wing Lian, and TheBloke for helpful advice
453
+ - And HUGE thanks to Wing Lian and the Axolotl contributors for making the best training framework!
454
+ - [<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)
455
+ - Thank you to all the other people in the Open Source AI community who have taught me and helped me along the way.
456
+
457
+ ## Example Output
458
+
459
+ Example output:
460
+ ```
461
+ <|im_start|>system
462
+ You are a helpful AI assistant<|im_end|>
463
+ <|im_start|>user
464
+ Which theory of everything is most likely correct?<|im_end|>
465
+ <|im_start|>assistant
466
+ The search for a "theory of everything" aims to unify all fundamental forces and particles in physics into one coherent framework. Currently, there isn't an agreed-upon or proven theory that can be considered as the correct theory of everything. However, two prominent theories have garnered significant attention:
467
+
468
+ 1. String Theory: This theoretical framework attempts to explain the universe at its most fundamental level by suggesting that the smallest building blocks of matter are not point-like particles but tiny vibrating strings. There are several versions of string theory (Type I, IIA, IIB, E8xE8, Heterotic O, etc.) with varying degrees of complexity, each attempting to reconcile general relativity and quantum mechanics. Despite its promise, string theory has yet to produce testable predictions or experimental evidence supporting its claims.
469
+
470
+ 2. Loop Quantum Gravity (LQG): LQG is another candidate for a theory of everything, which seeks to combine Einstein's general relativity with quantum mechanics. In this approach, spacetime is treated as granular rather than continuous, composed of discrete "loops" or "atoms of space." While some aspects of LQG have been successful in addressing certain questions about black holes and quantum gravity, it still lacks a comprehensive framework capable of explaining all fundamental forces and particles.
471
+
472
+ Both theories have their strengths and weaknesses, and neither has achieved widespread acceptance among physicists due to their lack of experimental support and inability to make definitive predictions. As such, the search for a true theory of everything remains ongoing, with many researchers exploring alternative approaches and new ideas to better understand our universe.
473
+ ```
474
+
475
+ [Buy me a coffee](https://www.buymeacoffee.com/ehartford)
476
+
477
+