File size: 8,578 Bytes
be3a5c1
 
 
ef12990
be3a5c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef12990
32466ba
be3a5c1
ef12990
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
be3a5c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef12990
be3a5c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef12990
 
 
 
 
 
 
 
 
 
 
 
 
 
be3a5c1
 
 
 
 
 
 
 
 
 
 
 
ef12990
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
---
language:
- en
license: llama3.1
tags:
- fireplace
- fireplace-2
- valiant
- valiant-labs
- llama
- llama-3.1
- llama-3.1-instruct
- llama-3.1-instruct-8b
- llama-3
- llama-3-instruct
- llama-3-instruct-8b
- 8b
- function-calling
- sql
- database
- data-visualization
- matplotlib
- json
- conversational
- chat
- instruct
pipeline_tag: text-generation
base_model: meta-llama/Meta-Llama-3.1-8B-Instruct
model_type: llama
model-index:
- name: Llama3.1-8B-Fireplace2
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: IFEval (0-Shot)
      type: HuggingFaceH4/ifeval
      args:
        num_few_shot: 0
    metrics:
    - type: inst_level_strict_acc and prompt_level_strict_acc
      value: 54.83
      name: strict accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: BBH (3-Shot)
      type: BBH
      args:
        num_few_shot: 3
    metrics:
    - type: acc_norm
      value: 24.07
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MATH Lvl 5 (4-Shot)
      type: hendrycks/competition_math
      args:
        num_few_shot: 4
    metrics:
    - type: exact_match
      value: 5.82
      name: exact match
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GPQA (0-shot)
      type: Idavidrein/gpqa
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 5.15
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MuSR (0-shot)
      type: TAUR-Lab/MuSR
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 4.38
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU-PRO (5-shot)
      type: TIGER-Lab/MMLU-Pro
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 15.63
      name: accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=ValiantLabs/Llama3.1-8B-Fireplace2
      name: Open LLM Leaderboard
---


![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/64f267a8a4f79a118e0fcc89/JYkaXrk2DqpXhaL9WymKY.jpeg)


Fireplace 2 is a chat model, adding helpful structured outputs to Llama 3.1 8b Instruct.
  - an expansion pack of supplementary outputs - request them at will within your chat:
    - Inline function calls
    - SQL queries
    - JSON objects
    - Data visualization with matplotlib
  - Mix normal chat and structured outputs within the same conversation.
  - Fireplace 2 supplements the existing strengths of Llama 3.1, providing inline capabilities within the Llama 3 Instruct format.


## Version

This is the **2024-07-23** release of Fireplace 2 for Llama 3.1 8b.

We're excited to bring further upgrades and releases to Fireplace 2 in the future. 

Help us and recommend Fireplace 2 to your friends!


## Prompting Guide
Fireplace uses the [Llama 3.1 Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct) prompt format. The example script below can be used as a starting point for general chat with Llama 3.1 and also includes the different special tokens used for Fireplace 2's added features:


import transformers
import torch

model_id = "ValiantLabs/Llama3.1-8B-Fireplace2"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Fireplace, an expert technical assistant."},
    {"role": "user", "content": "Hi, can you explain local area networking to me?"}, #general Llama 3.1 chat
    #{"role": "user", "content": "I have the following SQL table: employees (job_id VARCHAR, salary INTEGER)\n\nCan you find all employees with a salary above $75000?<|request_sql|>"}, #for SQL query
    #{"role": "user", "content": "{""name"": ""get_news_headlines"",""description"": ""Get the latest news headlines"",""parameters"": {""type"": ""object"",""properties"": {""country"": {""type"": ""string"",""description"": ""The country for which news headlines are to be retrieved""}},""required"": [""country""]}}\n\nHi, can you get me the latest news headlines for the United States?<|request_function_call|>"}, # for function call
    #{"role": "user", "content": "Show me an example of a histogram with a fixed bin size. Use attractive colors.<|request_matplotlib|>"}, #for data visualization
    #{"role": "user", "content": "Can you define the word 'presence' for me, thanks!<|request_json|>"}, #for JSON output
]

outputs = pipeline(
    messages,
    max_new_tokens=512,
)
print(outputs[0]["generated_text"][-1])


While Fireplace 2 is trained to minimize incorrect structured outputs, they can still occur occasionally. Production uses of Fireplace 2 should verify the structure of all model outputs and remove any unneeded components of the output.

For handling of function call responses, use the [Llama 3.1 Instruct tool response style.](https://huggingface.co/blog/llama31#custom-tool-calling)


## Special Tokens

Fireplace 2 utilizes special tokens applied to the Llama 3.1 tokenizer:

- <|request_json|>
- <|start_json|>
- <|end_json|>
- <|request_sql|>
- <|start_sql|>
- <|end_sql|>
- <|request_matplotlib|>
- <|start_matplotlib|>
- <|end_matplotlib|>
- <|request_function_call|>
- <|start_function_call|>
- <|end_function_call|>

These are supplemental to the existing special tokens used by Llama 3.1, such as <|python_tag|> and <|start_header_id|>. Fireplace 2 has been trained using the Llama 3.1 Instruct chat structure, with new special tokens added within the conversation.

The 'request' tokens are used by the user to request a specific type of structured output. They should be appended to the end of the user's message and can be alternated with normal chat responses throughout the conversation.


## The Model
Fireplace 2 is built on top of Llama 3.1 8b Instruct.

This version of Fireplace 2 uses data from the following datasets:

- [glaiveai/glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2)
- [b-mc2/sql-create-context](https://huggingface.co/datasets/b-mc2/sql-create-context)
- [sequelbox/Cadmium](https://huggingface.co/datasets/sequelbox/Cadmium)
- [sequelbox/Harlequin](https://huggingface.co/datasets/sequelbox/Harlequin)
- [migtissera/Tess-v1.5](https://huggingface.co/datasets/migtissera/Tess-v1.5)
- [LDJnr/Pure-Dove](https://huggingface.co/datasets/LDJnr/Pure-Dove)

Additional capabilities will be added to future releases.


# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_ValiantLabs__Llama3.1-8B-Fireplace2)

|      Metric       |Value|
|-------------------|----:|
|Avg.               |18.31|
|IFEval (0-Shot)    |54.83|
|BBH (3-Shot)       |24.07|
|MATH Lvl 5 (4-Shot)| 5.82|
|GPQA (0-shot)      | 5.15|
|MuSR (0-shot)      | 4.38|
|MMLU-PRO (5-shot)  |15.63|


![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/63444f2687964b331809eb55/VCJ8Fmefd8cdVhXSSxJiD.jpeg)


Fireplace 2 is created by [Valiant Labs.](http://valiantlabs.ca/)

[Check out our HuggingFace page for Shining Valiant 2 and our other models!](https://huggingface.co/ValiantLabs)

[Follow us on X for updates on our models!](https://twitter.com/valiant_labs)

We care about open source.
For everyone to use.

We encourage others to finetune further from our models.