adamo1139 commited on
Commit
0641089
1 Parent(s): 5bd8830

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +331 -3
README.md CHANGED
@@ -1,3 +1,331 @@
1
- ---
2
- license: llama3.1
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ license: llama3.1
5
+ tags:
6
+ - Llama-3
7
+ - instruct
8
+ - finetune
9
+ - chatml
10
+ - gpt4
11
+ - synthetic data
12
+ - distillation
13
+ - function calling
14
+ - json mode
15
+ - axolotl
16
+ - roleplaying
17
+ - chat
18
+ widget:
19
+ - example_title: Hermes 3
20
+ messages:
21
+ - role: system
22
+ content: >-
23
+ You are a sentient, superintelligent artificial general intelligence, here
24
+ to teach and assist me.
25
+ - role: user
26
+ content: What is the meaning of life?
27
+ model-index:
28
+ - name: Hermes-3-Llama-3.1-70B
29
+ results: []
30
+ base_model:
31
+ - NousResearch/Hermes-3-Llama-3.1-8B
32
+ ---
33
+ # Hermes 3 - Llama-3.1 8B FP8
34
+
35
+ <img src="https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/bMcZ3sNNQK8SRZpHXBmwM.jpeg" width="500" style="float:center">
36
+
37
+ ## Model Description
38
+
39
+ This is a FP8 Dynamic version of the model.
40
+
41
+ Hermes 3 is the latest version of our flagship Hermes series of LLMs by Nous Research.
42
+
43
+ For more details on new capabilities, training results, and more, see the [**Hermes 3 Technical Report**](https://arxiv.org/abs/2408.11857).
44
+
45
+ Hermes 3 is a generalist language model with many improvements over Hermes 2, including advanced agentic capabilities, much better roleplaying, reasoning, multi-turn conversation, long context coherence, and improvements across the board.
46
+
47
+ The ethos of the Hermes series of models is focused on aligning LLMs to the user, with powerful steering capabilities and control given to the end user.
48
+
49
+ The Hermes 3 series builds and expands on the Hermes 2 set of capabilities, including more powerful and reliable function calling and structured output capabilities, generalist assistant capabilities, and improved code generation skills.
50
+
51
+ ## Creation
52
+
53
+ This quant was created using llmcompressor.
54
+
55
+ Code below.
56
+
57
+ ```python
58
+ import torch
59
+ from datasets import load_dataset
60
+ from transformers import AutoTokenizer
61
+
62
+ from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
63
+ from llmcompressor.transformers.compression.helpers import (
64
+ calculate_offload_device_map,
65
+ custom_offload_device_map,
66
+ )
67
+
68
+ recipe = """
69
+ quant_stage:
70
+ quant_modifiers:
71
+ QuantizationModifier:
72
+ ignore: ["lm_head"]
73
+ config_groups:
74
+ group_0:
75
+ weights:
76
+ num_bits: 8
77
+ type: float
78
+ strategy: tensor
79
+ dynamic: false
80
+ symmetric: true
81
+ input_activations:
82
+ num_bits: 8
83
+ type: float
84
+ strategy: tensor
85
+ dynamic: false
86
+ symmetric: true
87
+ targets: ["Linear"]
88
+ """
89
+
90
+ model_stub = "NousResearch/Hermes-3-Llama-3.1-8B"
91
+ model_name = model_stub.split("/")[-1]
92
+
93
+ device_map = calculate_offload_device_map(
94
+ model_stub, reserve_for_hessians=False, num_gpus=1, torch_dtype="auto"
95
+ )
96
+
97
+ model = SparseAutoModelForCausalLM.from_pretrained(
98
+ model_stub, torch_dtype="auto", device_map=device_map
99
+ )
100
+ tokenizer = AutoTokenizer.from_pretrained(model_stub)
101
+
102
+ output_dir = f"./{model_name}-FP8"
103
+
104
+ DATASET_ID = "HuggingFaceH4/ultrachat_200k"
105
+ DATASET_SPLIT = "train_sft"
106
+ NUM_CALIBRATION_SAMPLES = 512
107
+ MAX_SEQUENCE_LENGTH = 4096
108
+
109
+ ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
110
+ ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))
111
+
112
+ def preprocess(example):
113
+ return {
114
+ "text": tokenizer.apply_chat_template(
115
+ example["messages"],
116
+ tokenize=False,
117
+ )
118
+ }
119
+
120
+ ds = ds.map(preprocess)
121
+
122
+ def tokenize(sample):
123
+ return tokenizer(
124
+ sample["text"],
125
+ padding=False,
126
+ max_length=MAX_SEQUENCE_LENGTH,
127
+ truncation=True,
128
+ add_special_tokens=False,
129
+ )
130
+
131
+ ds = ds.map(tokenize, remove_columns=ds.column_names)
132
+
133
+ oneshot(
134
+ model=model,
135
+ output_dir=output_dir,
136
+ dataset=ds,
137
+ recipe=recipe,
138
+ max_seq_length=MAX_SEQUENCE_LENGTH,
139
+ num_calibration_samples=NUM_CALIBRATION_SAMPLES,
140
+ save_compressed=True,
141
+ )
142
+
143
+ ```
144
+
145
+
146
+ # Benchmarks
147
+
148
+ Hermes 3 is competitive, if not superior, to Llama-3.1 Instruct models at general capabilities, with varying strengths and weaknesses attributable between the two.
149
+
150
+ Full benchmark comparisons below:
151
+
152
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/DIMca3M0U-ArWwtyIbF-k.png)
153
+
154
+
155
+ # Prompt Format
156
+
157
+ Hermes 3 uses ChatML as the prompt format, opening up a much more structured system for engaging the LLM in multi-turn chat dialogue.
158
+
159
+ System prompts allow steerability and interesting new ways to interact with an LLM, guiding rules, roles, and stylistic choices of the model.
160
+
161
+ 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.
162
+
163
+ 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.
164
+
165
+ Prompt with system instruction (Use whatever system prompt you like, this is just an example!):
166
+ ```
167
+ <|im_start|>system
168
+ You are Hermes 3, 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|>
169
+ <|im_start|>user
170
+ Hello, who are you?<|im_end|>
171
+ <|im_start|>assistant
172
+ Hi there! My name is Hermes 3, 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|>
173
+ ```
174
+
175
+ This prompt is available as a [chat template](https://huggingface.co/docs/transformers/main/chat_templating), which means you can format messages using the
176
+ `tokenizer.apply_chat_template()` method:
177
+
178
+ ```python
179
+ messages = [
180
+ {"role": "system", "content": "You are Hermes 3."},
181
+ {"role": "user", "content": "Hello, who are you?"}
182
+ ]
183
+ gen_input = tokenizer.apply_chat_template(messages, return_tensors="pt")
184
+ model.generate(**gen_input)
185
+ ```
186
+
187
+ 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
188
+ that the model continues with an assistant response.
189
+
190
+ To utilize the prompt format without a system prompt, simply leave the line out.
191
+
192
+
193
+ ## Prompt Format for Function Calling
194
+
195
+ Our model was trained on specific system prompts and structures for Function Calling.
196
+
197
+ You should use the system role with this message, followed by a function signature json as this example shows here.
198
+ ```
199
+ <|im_start|>system
200
+ 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:
201
+ <tool_call>
202
+ {"arguments": <args-dict>, "name": <function-name>}
203
+ </tool_call><|im_end|>
204
+ ```
205
+
206
+ To complete the function call, create a user prompt that follows the above system prompt, like so:
207
+ ```
208
+ <|im_start|>user
209
+ Fetch the stock fundamentals data for Tesla (TSLA)<|im_end|>
210
+ ```
211
+
212
+ 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):
213
+ ```
214
+ <|im_start|>assistant
215
+ <tool_call>
216
+ {"arguments": {"symbol": "TSLA"}, "name": "get_stock_fundamentals"}
217
+ </tool_call><|im_end|>
218
+ ```
219
+
220
+ 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:
221
+ ```
222
+ <|im_start|>tool
223
+ <tool_response>
224
+ {"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}}
225
+ </tool_response>
226
+ <|im_end|>
227
+ ```
228
+
229
+ The assistant will then read in that data from the function's response, and generate a natural language response:
230
+ ```
231
+ <|im_start|>assistant
232
+ The stock fundamentals data for Tesla (TSLA) are as follows:
233
+ - **Symbol**: TSLA
234
+ - **Company Name**: Tesla, Inc.
235
+ - **Sector**: Consumer Cyclical
236
+ - **Industry**: Auto Manufacturers
237
+ - **Market Capitalization**: $566,160,130,480
238
+ - **Forward Price-to-Earnings Ratio (PE Ratio)**: 42.73
239
+ - **Price-to-Book Ratio (PB Ratio)**: 9.04
240
+ - **Dividend Yield**: N/A
241
+ - **Trailing Earnings Per Share (EPS)**: $4.3
242
+ - **Beta Value of the Stock**: 2.42
243
+ - **52-Week High Price of the Stock**: $299.29
244
+ - **52-Week Low Price of the Stock**: $152.37
245
+
246
+ 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|>
247
+ ```
248
+
249
+
250
+ ## Prompt Format for JSON Mode / Structured Outputs
251
+
252
+ 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.
253
+
254
+ 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
255
+
256
+ ```
257
+ <|im_start|>system
258
+ 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|>
259
+ ```
260
+
261
+ 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.
262
+
263
+
264
+ # Inference
265
+
266
+ Here is example code using HuggingFace Transformers to inference the model
267
+
268
+ ```python
269
+ # Code to inference Hermes with HF Transformers
270
+ # Requires pytorch, transformers, bitsandbytes, sentencepiece, protobuf, and flash-attn packages
271
+
272
+ import torch
273
+ from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaForCausalLM
274
+ import bitsandbytes, flash_attn
275
+
276
+ tokenizer = AutoTokenizer.from_pretrained('NousResearch/Hermes-3-Llama-3.1-8B', trust_remote_code=True)
277
+ model = LlamaForCausalLM.from_pretrained(
278
+ "NousResearch/Hermes-3-Llama-3.1-8B",
279
+ torch_dtype=torch.float16,
280
+ device_map="auto",
281
+ load_in_8bit=False,
282
+ load_in_4bit=True,
283
+ use_flash_attention_2=True
284
+ )
285
+
286
+ prompts = [
287
+ """<|im_start|>system
288
+ You are a sentient, superintelligent artificial general intelligence, here to teach and assist me.<|im_end|>
289
+ <|im_start|>user
290
+ Write a short story about Goku discovering kirby has teamed up with Majin Buu to destroy the world.<|im_end|>
291
+ <|im_start|>assistant""",
292
+ ]
293
+
294
+ for chat in prompts:
295
+ print(chat)
296
+ input_ids = tokenizer(chat, return_tensors="pt").input_ids.to("cuda")
297
+ 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)
298
+ response = tokenizer.decode(generated_ids[0][input_ids.shape[-1]:], skip_special_tokens=True, clean_up_tokenization_space=True)
299
+ print(f"Response: {response}")
300
+ ```
301
+
302
+ You can also run this model with vLLM, by running the following in your terminal after `pip install vllm`
303
+
304
+ `vllm serve NousResearch/Hermes-3-Llama-3.1-8B`
305
+
306
+ ## Inference Code for Function Calling:
307
+
308
+ All code for utilizing, parsing, and building function calling templates is available on our github:
309
+ [https://github.com/NousResearch/Hermes-Function-Calling](https://github.com/NousResearch/Hermes-Function-Calling)
310
+
311
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/6317aade83d8d2fd903192d9/oi4CiGh50xmoviUQnh8R3.png)
312
+
313
+
314
+ ## Quantized Versions:
315
+
316
+ GGUF Quants: https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B-GGUF
317
+ FP8 Quants: https://huggingface.co/adamo1139/Hermes-3-Llama-3.1-8B-FP8
318
+
319
+ # How to cite:
320
+
321
+ ```bibtext
322
+ @misc{teknium2024hermes3technicalreport,
323
+ title={Hermes 3 Technical Report},
324
+ author={Ryan Teknium and Jeffrey Quesnelle and Chen Guang},
325
+ year={2024},
326
+ eprint={2408.11857},
327
+ archivePrefix={arXiv},
328
+ primaryClass={cs.CL},
329
+ url={https://arxiv.org/abs/2408.11857},
330
+ }
331
+ ```