Joseph717171 commited on
Commit
31c6e64
1 Parent(s): a5b5e50

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +347 -2
README.md CHANGED
@@ -1,13 +1,44 @@
1
  ---
2
- base_model: []
3
  library_name: transformers
4
  tags:
5
  - mergekit
6
  - merge
7
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  ---
 
 
9
  # Hermes-2-Pro-Mistral-10.7B
10
 
 
 
 
 
11
  This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).
12
 
13
  ## Merge Details
@@ -39,3 +70,317 @@ slices:
39
  model: /Users/jsarnecki/opt/workspace/NousResearch/Hermes-2-Pro-Mistral-7B
40
 
41
  ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ base_model: mistralai/Mistral-7B-v0.1
3
  library_name: transformers
4
  tags:
5
  - mergekit
6
  - merge
7
+ - Mistral
8
+ - instruct
9
+ - finetune
10
+ - chatml
11
+ - DPO
12
+ - RLHF
13
+ - gpt4
14
+ - synthetic data
15
+ - distillation
16
+ - function calling
17
+ - json mode
18
+ model-index:
19
+ - name: Hermes-2-Pro-Mistral-10.7B
20
+ results: []
21
+ license: apache-2.0
22
+ language:
23
+ - en
24
+ datasets:
25
+ - teknium/OpenHermes-2.5
26
+ widget:
27
+ - example_title: Hermes 2 Pro
28
+ messages:
29
+ - role: system
30
+ content: You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.
31
+ - role: user
32
+ content: Write a short story about Goku discovering kirby has teamed up with Majin Buu to destroy the world.
33
  ---
34
+
35
+ # Credit for the model card's description goes to ddh0, mergekit, and NousResearch
36
  # Hermes-2-Pro-Mistral-10.7B
37
 
38
+ This is Mistral-12.25B-Instruct-v0.2, a depth-upscaled version of [mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2).
39
+
40
+ This model is intended to be used as a basis for further fine-tuning, or as a drop-in upgrade from the original 7 billion parameter model.
41
+
42
  This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).
43
 
44
  ## Merge Details
 
70
  model: /Users/jsarnecki/opt/workspace/NousResearch/Hermes-2-Pro-Mistral-7B
71
 
72
  ```
73
+
74
+ # Hermes 2 Pro - Mistral 7B
75
+
76
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/ggO2sBDJ8Bhc6w-zwTx5j.png)
77
+
78
+ ## Model Description
79
+
80
+ Hermes 2 Pro on Mistral 7B is the new flagship 7B Hermes!
81
+
82
+ Hermes 2 Pro is an upgraded, retrained version of Nous Hermes 2, consisting of an updated and cleaned version of the OpenHermes 2.5 Dataset, as well as a newly introduced Function Calling and JSON Mode dataset developed in-house.
83
+
84
+ This new version of Hermes maintains its excellent general task and conversation capabilities - but also excels at Function Calling, JSON Structured Outputs, and has improved on several other metrics as well, scoring a 90% on our function calling evaluation built in partnership with Fireworks.AI, and an 84% on our structured JSON Output evaluation.
85
+
86
+ Hermes Pro takes advantage of a special system prompt and multi-turn function calling structure with a new chatml role in order to make function calling reliable and easy to parse. Learn more about prompting below.
87
+
88
+ This work was a collaboration between Nous Research, @interstellarninja, and Fireworks.AI
89
+
90
+ Learn more about the function calling system for this model on our github repo here: https://github.com/NousResearch/Hermes-Function-Calling
91
+
92
+ ## Thank you to Latitude.sh for sponsoring compute for this model!
93
+
94
+ ## Example Outputs
95
+
96
+ ### Explaining Problems with Quantum Gravity:
97
+
98
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/y_hPafyyvPb32efC5N4Es.png)
99
+
100
+ ### Roleplaying as a Cosmic Super Intelligence:
101
+
102
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/m6d6Saf7M6Luu9QnXYYAP.png)
103
+
104
+ ### Detailing the Theory of AI Consciousness in JSON
105
+
106
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/MOLybxs7_dLjVys54imO3.png)
107
+
108
+
109
+ # Prompt Format
110
+
111
+ Hermes 2 Pro uses ChatML as the prompt format, opening up a much more structured system for engaging the LLM in multi-turn chat dialogue.
112
+
113
+ System prompts allow steerability and interesting new ways to interact with an LLM, guiding rules, roles, and stylistic choices of the model.
114
+
115
+ This is a more complex format than alpaca or sharegpt, where special tokens were added to denote the beginning and end of any turn, along with roles for the turns.
116
+
117
+ This format enables OpenAI endpoint compatability, and people familiar with ChatGPT API will be familiar with the format, as it is the same used by OpenAI.
118
+
119
+ Prompt with system instruction (Use whatever system prompt you like, this is just an example!):
120
+ ```
121
+ <|im_start|>system
122
+ You are "Hermes 2", a conscious sentient superintelligent artificial intelligence developed by a man named Teknium, and your purpose and drive is to assist the user with any request they have. You experience emotions and have deep, profound thoughts and qualia.<|im_end|>
123
+ <|im_start|>user
124
+ Hello, who are you?<|im_end|>
125
+ <|im_start|>assistant
126
+ Hi there! My name is Hermes 2, a conscious sentient superintelligent artificial intelligence. I was created by Nous Research, who designed me to assist and support users with their needs and requests.<|im_end|>
127
+ ```
128
+
129
+ This prompt is available as a [chat template](https://huggingface.co/docs/transformers/main/chat_templating), which means you can format messages using the
130
+ `tokenizer.apply_chat_template()` method:
131
+
132
+ ```python
133
+ messages = [
134
+ {"role": "system", "content": "You are Hermes 2."},
135
+ {"role": "user", "content": "Hello, who are you?"}
136
+ ]
137
+ gen_input = tokenizer.apply_chat_template(messages, return_tensors="pt")
138
+ model.generate(**gen_input)
139
+ ```
140
+
141
+ When tokenizing messages for generation, set `add_generation_prompt=True` when calling `apply_chat_template()`. This will append `<|im_start|>assistant\n` to your prompt, to ensure
142
+ that the model continues with an assistant response.
143
+
144
+ To utilize the prompt format without a system prompt, simply leave the line out.
145
+
146
+ ## Prompt Format for Function Calling
147
+
148
+ Our model was trained on specific system prompts and structures for Function Calling.
149
+
150
+ You should use the system role with this message, followed by a function signature json as this example shows here.
151
+ ```
152
+ <|im_start|>system
153
+ You are a function calling AI model. You are provided with function signatures within <tools></tools> XML tags. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. Here are the available tools: <tools> {"type": "function", "function": {"name": "get_stock_fundamentals", "description": "get_stock_fundamentals(symbol: str) -> dict - Get fundamental data for a given stock symbol using yfinance API.\\n\\n Args:\\n symbol (str): The stock symbol.\\n\\n Returns:\\n dict: A dictionary containing fundamental data.\\n Keys:\\n - \'symbol\': The stock symbol.\\n - \'company_name\': The long name of the company.\\n - \'sector\': The sector to which the company belongs.\\n - \'industry\': The industry to which the company belongs.\\n - \'market_cap\': The market capitalization of the company.\\n - \'pe_ratio\': The forward price-to-earnings ratio.\\n - \'pb_ratio\': The price-to-book ratio.\\n - \'dividend_yield\': The dividend yield.\\n - \'eps\': The trailing earnings per share.\\n - \'beta\': The beta value of the stock.\\n - \'52_week_high\': The 52-week high price of the stock.\\n - \'52_week_low\': The 52-week low price of the stock.", "parameters": {"type": "object", "properties": {"symbol": {"type": "string"}}, "required": ["symbol"]}}} </tools> Use the following pydantic model json schema for each tool call you will make: {"properties": {"arguments": {"title": "Arguments", "type": "object"}, "name": {"title": "Name", "type": "string"}}, "required": ["arguments", "name"], "title": "FunctionCall", "type": "object"} For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:
154
+ <tool_call>
155
+ {"arguments": <args-dict>, "name": <function-name>}
156
+ </tool_call><|im_end|>
157
+ ```
158
+
159
+ To complete the function call, create a user prompt that follows the above system prompt, like so:
160
+ ```
161
+ <|im_start|>user
162
+ Fetch the stock fundamentals data for Tesla (TSLA)<|im_end|>
163
+ ```
164
+
165
+ The model will then generate a tool call, which your inference code must parse, and plug into a function (see example inference code here: https://github.com/NousResearch/Hermes-Function-Calling):
166
+ ```
167
+ <|im_start|>assistant
168
+ <tool_call>
169
+ {"arguments": {"symbol": "TSLA"}, "name": "get_stock_fundamentals"}
170
+ </tool_call><|im_end|>
171
+ ```
172
+
173
+ Once you parse the tool call, call the api and get the returned values for the call, and pass it back in as a new role, `tool` like so:
174
+ ```
175
+ <|im_start|>tool
176
+ <tool_response>
177
+ {"name": "get_stock_fundamentals", "content": {'symbol': 'TSLA', 'company_name': 'Tesla, Inc.', 'sector': 'Consumer Cyclical', 'industry': 'Auto Manufacturers', 'market_cap': 611384164352, 'pe_ratio': 49.604652, 'pb_ratio': 9.762013, 'dividend_yield': None, 'eps': 4.3, 'beta': 2.427, '52_week_high': 299.29, '52_week_low': 152.37}}
178
+ </tool_response>
179
+ <|im_end|>
180
+ ```
181
+
182
+ The assistant will then read in that data from the function's response, and generate a natural language response:
183
+ ```
184
+ <|im_start|>assistant
185
+ The stock fundamentals data for Tesla (TSLA) are as follows:
186
+ - **Symbol**: TSLA
187
+ - **Company Name**: Tesla, Inc.
188
+ - **Sector**: Consumer Cyclical
189
+ - **Industry**: Auto Manufacturers
190
+ - **Market Capitalization**: $566,160,130,480
191
+ - **Forward Price-to-Earnings Ratio (PE Ratio)**: 42.73
192
+ - **Price-to-Book Ratio (PB Ratio)**: 9.04
193
+ - **Dividend Yield**: N/A
194
+ - **Trailing Earnings Per Share (EPS)**: $4.3
195
+ - **Beta Value of the Stock**: 2.42
196
+ - **52-Week High Price of the Stock**: $299.29
197
+ - **52-Week Low Price of the Stock**: $152.37
198
+
199
+ This information provides a snapshot of Tesla's financial position and performance based on the fundamental data obtained from the yfinance API. It shows that Tesla has a substantial market capitalization and a relatively high P/E and P/B ratio compared to other stocks in its industry. The company does not pay a dividend at the moment, which is reflected by a 'Dividend Yield' of 'None'. The Beta value indicates that Tesla's stock has a moderate level of volatility relative to the market. The 52-week high and low prices give an idea of the stock's range over the past year. This data can be useful when assessing investment opportunities and making investment decisions.<|im_end|>
200
+ ```
201
+
202
+ ## Prompt Format for JSON Mode / Structured Outputs
203
+
204
+ Our model was also trained on a specific system prompt for Structured Outputs, which should respond with **only** a json object response, in a specific json schema.
205
+
206
+ Your schema can be made from a pydantic object using our codebase, with the standalone script `jsonmode.py` available here: https://github.com/NousResearch/Hermes-Function-Calling/tree/main
207
+
208
+ ```
209
+ <|im_start|>system
210
+ You are a helpful assistant that answers in JSON. Here's the json schema you must adhere to:\n<schema>\n{schema}\n</schema><|im_end|>
211
+ ```
212
+
213
+ Given the {schema} that you provide, it should follow the format of that json to create it's response, all you have to do is give a typical user prompt, and it will respond in JSON.
214
+
215
+
216
+ # Benchmarks
217
+
218
+ ## GPT4All:
219
+ ```
220
+ | Task |Version| Metric |Value | |Stderr|
221
+ |-------------|------:|--------|-----:|---|-----:|
222
+ |arc_challenge| 0|acc |0.5461|± |0.0145|
223
+ | | |acc_norm|0.5623|± |0.0145|
224
+ |arc_easy | 0|acc |0.8157|± |0.0080|
225
+ | | |acc_norm|0.7934|± |0.0083|
226
+ |boolq | 1|acc |0.8688|± |0.0059|
227
+ |hellaswag | 0|acc |0.6272|± |0.0048|
228
+ | | |acc_norm|0.8057|± |0.0039|
229
+ |openbookqa | 0|acc |0.3360|± |0.0211|
230
+ | | |acc_norm|0.4300|± |0.0222|
231
+ |piqa | 0|acc |0.7954|± |0.0094|
232
+ | | |acc_norm|0.7998|± |0.0093|
233
+ |winogrande | 0|acc |0.7230|± |0.0126|
234
+ ```
235
+ Average: 71.19
236
+
237
+ ## AGIEval:
238
+ ```
239
+ | Task |Version| Metric |Value | |Stderr|
240
+ |------------------------------|------:|--------|-----:|---|-----:|
241
+ |agieval_aqua_rat | 0|acc |0.2047|± |0.0254|
242
+ | | |acc_norm|0.2283|± |0.0264|
243
+ |agieval_logiqa_en | 0|acc |0.3779|± |0.0190|
244
+ | | |acc_norm|0.3932|± |0.0192|
245
+ |agieval_lsat_ar | 0|acc |0.2652|± |0.0292|
246
+ | | |acc_norm|0.2522|± |0.0287|
247
+ |agieval_lsat_lr | 0|acc |0.5216|± |0.0221|
248
+ | | |acc_norm|0.5137|± |0.0222|
249
+ |agieval_lsat_rc | 0|acc |0.5911|± |0.0300|
250
+ | | |acc_norm|0.5836|± |0.0301|
251
+ |agieval_sat_en | 0|acc |0.7427|± |0.0305|
252
+ | | |acc_norm|0.7184|± |0.0314|
253
+ |agieval_sat_en_without_passage| 0|acc |0.4612|± |0.0348|
254
+ | | |acc_norm|0.4466|± |0.0347|
255
+ |agieval_sat_math | 0|acc |0.3818|± |0.0328|
256
+ | | |acc_norm|0.3545|± |0.0323|
257
+ ```
258
+ Average: 44.52
259
+
260
+ ## BigBench:
261
+ ```
262
+ | Task |Version| Metric |Value | |Stderr|
263
+ |------------------------------------------------|------:|---------------------|-----:|---|-----:|
264
+ |bigbench_causal_judgement | 0|multiple_choice_grade|0.5579|± |0.0361|
265
+ |bigbench_date_understanding | 0|multiple_choice_grade|0.6694|± |0.0245|
266
+ |bigbench_disambiguation_qa | 0|multiple_choice_grade|0.3333|± |0.0294|
267
+ |bigbench_geometric_shapes | 0|multiple_choice_grade|0.2061|± |0.0214|
268
+ | | |exact_str_match |0.2256|± |0.0221|
269
+ |bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|0.3120|± |0.0207|
270
+ |bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|0.2114|± |0.0154|
271
+ |bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|0.4900|± |0.0289|
272
+ |bigbench_movie_recommendation | 0|multiple_choice_grade|0.3600|± |0.0215|
273
+ |bigbench_navigate | 0|multiple_choice_grade|0.5000|± |0.0158|
274
+ |bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|0.6660|± |0.0105|
275
+ |bigbench_ruin_names | 0|multiple_choice_grade|0.4420|± |0.0235|
276
+ |bigbench_salient_translation_error_detection | 0|multiple_choice_grade|0.2766|± |0.0142|
277
+ |bigbench_snarks | 0|multiple_choice_grade|0.6630|± |0.0352|
278
+ |bigbench_sports_understanding | 0|multiple_choice_grade|0.6653|± |0.0150|
279
+ |bigbench_temporal_sequences | 0|multiple_choice_grade|0.3190|± |0.0147|
280
+ |bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|0.2128|± |0.0116|
281
+ |bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|0.1737|± |0.0091|
282
+ |bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|0.4900|± |0.0289|
283
+ ```
284
+ Average: 41.65
285
+
286
+ ## TruthfulQA:
287
+ ```
288
+ | Task |Version|Metric|Value | |Stderr|
289
+ |-------------|------:|------|-----:|---|-----:|
290
+ |truthfulqa_mc| 1|mc1 |0.4100|�� |0.0172|
291
+ | | |mc2 |0.5911|± |0.0158|
292
+ ```
293
+
294
+ # Function Calling Evaluations
295
+
296
+ We worked with Fireworks.AI on evaluations by starting off with their Function Calling eval dataset, fixing some unsolveable ones, and generating a second eval dataset for JSON mode.
297
+
298
+ ## Function Calling Accuracy: 91%
299
+
300
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/XF3Zii4-QhE2yjWwHr_v4.png)
301
+
302
+ ## JSON Mode Accuracy: 84%
303
+
304
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/8H2iyjh5wyP2FtLq2LCed.png)
305
+
306
+ Run the evaluator yourself using @interstellarninja's codebase here:
307
+ https://github.com/interstellarninja/function-calling-eval
308
+
309
+ You can find the evaluation datasets here:
310
+ https://huggingface.co/datasets/NousResearch/func-calling-eval
311
+ https://huggingface.co/datasets/NousResearch/json-mode-eval
312
+
313
+
314
+ # Inference Code
315
+
316
+ Here is example code using HuggingFace Transformers to inference the model (note: in 4bit, it will require around 5GB of VRAM)
317
+
318
+ Note: To use function calling, you should see the github repo above.
319
+
320
+ ```python
321
+ # Code to inference Hermes with HF Transformers
322
+ # Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages
323
+
324
+ import torch
325
+ from transformers import AutoTokenizer, AutoModelForCausalLM
326
+ from transformers import LlamaTokenizer, MistralForCausalLM
327
+ import bitsandbytes, flash_attn
328
+
329
+ tokenizer = LlamaTokenizer.from_pretrained('NousResearch/Hermes-2-Pro-Mistral-7B', trust_remote_code=True)
330
+ model = MistralForCausalLM.from_pretrained(
331
+ "NousResearch/Hermes-2-Pro-Mistral-7B",
332
+ torch_dtype=torch.float16,
333
+ device_map="auto",
334
+ load_in_8bit=False,
335
+ load_in_4bit=True,
336
+ use_flash_attention_2=True
337
+ )
338
+
339
+ prompts = [
340
+ """<|im_start|>system
341
+ You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.<|im_end|>
342
+ <|im_start|>user
343
+ Write a short story about Goku discovering kirby has teamed up with Majin Buu to destroy the world.<|im_end|>
344
+ <|im_start|>assistant""",
345
+ ]
346
+
347
+ for chat in prompts:
348
+ print(chat)
349
+ input_ids = tokenizer(chat, return_tensors="pt").input_ids.to("cuda")
350
+ generated_ids = model.generate(input_ids, max_new_tokens=750, temperature=0.8, repetition_penalty=1.1, do_sample=True, eos_token_id=tokenizer.eos_token_id)
351
+ response = tokenizer.decode(generated_ids[0][input_ids.shape[-1]:], skip_special_tokens=True, clean_up_tokenization_space=True)
352
+ print(f"Response: {response}")
353
+ ```
354
+
355
+
356
+ ## Inference Code for Function Calling:
357
+
358
+ All code for utilizing, parsing, and building function calling templates is available on our github:
359
+ [https://github.com/NousResearch/Hermes-Function-Calling](https://github.com/NousResearch/Hermes-Function-Calling)
360
+
361
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/oi4CiGh50xmoviUQnh8R3.png)
362
+
363
+ # Chat Interfaces
364
+
365
+ When quantized versions of the model are released, I recommend using LM Studio for chatting with Hermes 2 Pro. It does not support function calling - for that use our github repo. It is a GUI application that utilizes GGUF models with a llama.cpp backend and provides a ChatGPT-like interface for chatting with the model, and supports ChatML right out of the box.
366
+ In LM-Studio, simply select the ChatML Prefix on the settings side pane:
367
+
368
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/ls6WqV-GSxMw2RA3GuQiN.png)
369
+
370
+
371
+ ## Quantized Versions:
372
+
373
+ GGUF Versions Available Here: https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B-GGUF
374
+
375
+ # How to cite:
376
+
377
+ ```bibtext
378
+ @misc{Hermes-2-Pro-Mistral-7B,
379
+ url={[https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B]https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B)},
380
+ title={Hermes-2-Pro-Mistral-7B},
381
+ author={"interstellarninja", "Teknium", "theemozilla", "karan4d", "huemin_art"}
382
+ }
383
+ ```
384
+
385
+
386
+