Update README.md
Browse files
README.md
CHANGED
@@ -72,8 +72,182 @@ You can use the model with the Hugging Face `transformers` and the rubra library
|
|
72 |
pip install rubra_tools torch==2.3.0 transformers
|
73 |
```
|
74 |
|
|
|
75 |
```python
|
76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
```
|
78 |
|
79 |
## Training Hyperparameters
|
|
|
72 |
pip install rubra_tools torch==2.3.0 transformers
|
73 |
```
|
74 |
|
75 |
+
### 1. Load the Model
|
76 |
```python
|
77 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
78 |
+
import torch
|
79 |
+
from rubra_tools import preprocess_input, postprocess_output
|
80 |
+
|
81 |
+
model_id = "rubra-ai/Meta-Llama-3-8B-Instruct"
|
82 |
+
|
83 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
84 |
+
model = AutoModelForCausalLM.from_pretrained(
|
85 |
+
model_id,
|
86 |
+
torch_dtype=torch.bfloat16,
|
87 |
+
device_map="auto",
|
88 |
+
)
|
89 |
+
```
|
90 |
+
|
91 |
+
### 2. Define Functions
|
92 |
+
|
93 |
+
Here we use 4 functions for a simple math chaining question:
|
94 |
+
```python
|
95 |
+
functions = [
|
96 |
+
{
|
97 |
+
'type': 'function',
|
98 |
+
'function': {
|
99 |
+
'name': 'addition',
|
100 |
+
'description': "Adds two numbers together",
|
101 |
+
'parameters': {
|
102 |
+
'type': 'object',
|
103 |
+
'properties': {
|
104 |
+
'a': {
|
105 |
+
'description': 'First number to add',
|
106 |
+
'type': 'string'
|
107 |
+
},
|
108 |
+
'b': {
|
109 |
+
'description': 'Second number to add',
|
110 |
+
'type': 'string'
|
111 |
+
}
|
112 |
+
},
|
113 |
+
'required': []
|
114 |
+
}
|
115 |
+
}
|
116 |
+
},
|
117 |
+
{
|
118 |
+
'type': 'function',
|
119 |
+
'function': {
|
120 |
+
'name': 'subtraction',
|
121 |
+
'description': "Subtracts two numbers",
|
122 |
+
'parameters': {
|
123 |
+
'type': 'object',
|
124 |
+
'properties': {
|
125 |
+
'a': {
|
126 |
+
'description': 'First number to be subtracted from',
|
127 |
+
'type': 'string'
|
128 |
+
},
|
129 |
+
'b': {
|
130 |
+
'description': 'Number to subtract',
|
131 |
+
'type': 'string'
|
132 |
+
}
|
133 |
+
},
|
134 |
+
'required': []
|
135 |
+
}
|
136 |
+
}
|
137 |
+
},
|
138 |
+
{
|
139 |
+
'type': 'function',
|
140 |
+
'function': {
|
141 |
+
'name': 'multiplication',
|
142 |
+
'description': "Multiply two numbers together",
|
143 |
+
'parameters': {
|
144 |
+
'type': 'object',
|
145 |
+
'properties': {
|
146 |
+
'a': {
|
147 |
+
'description': 'First number to multiply',
|
148 |
+
'type': 'string'
|
149 |
+
},
|
150 |
+
'b': {
|
151 |
+
'description': 'Second number to multiply',
|
152 |
+
'type': 'string'
|
153 |
+
}
|
154 |
+
},
|
155 |
+
'required': []
|
156 |
+
}
|
157 |
+
}
|
158 |
+
},
|
159 |
+
{
|
160 |
+
'type': 'function',
|
161 |
+
'function': {
|
162 |
+
'name': 'division',
|
163 |
+
'description': "Divide two numbers",
|
164 |
+
'parameters': {
|
165 |
+
'type': 'object',
|
166 |
+
'properties': {
|
167 |
+
'a': {
|
168 |
+
'description': 'First number to use as the dividend',
|
169 |
+
'type': 'string'
|
170 |
+
},
|
171 |
+
'b': {
|
172 |
+
'description': 'Second number to use as the divisor',
|
173 |
+
'type': 'string'
|
174 |
+
}
|
175 |
+
},
|
176 |
+
'required': []
|
177 |
+
}
|
178 |
+
}
|
179 |
+
},
|
180 |
+
]
|
181 |
+
```
|
182 |
+
|
183 |
+
### 3. Start the conversation
|
184 |
+
```python
|
185 |
+
messages = [
|
186 |
+
{"role": "system", "content": "You are a helpful assistant."},
|
187 |
+
{"role": "user", "content": "What is the result of four plus six? Take the result and add 2? Then multiply by 5 and then divide by two"},
|
188 |
+
]
|
189 |
+
|
190 |
+
def run_model(messages, functions):
|
191 |
+
## Format messages in Rubra's format
|
192 |
+
formatted_msgs = preprocess_input(msgs=messages, tools=functions)
|
193 |
+
|
194 |
+
input_ids = tokenizer.apply_chat_template(
|
195 |
+
formatted_msgs,
|
196 |
+
add_generation_prompt=True,
|
197 |
+
return_tensors="pt"
|
198 |
+
).to(model.device)
|
199 |
+
|
200 |
+
terminators = [
|
201 |
+
tokenizer.eos_token_id,
|
202 |
+
tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
203 |
+
]
|
204 |
+
|
205 |
+
outputs = model.generate(
|
206 |
+
input_ids,
|
207 |
+
max_new_tokens=1000,
|
208 |
+
eos_token_id=terminators,
|
209 |
+
do_sample=True,
|
210 |
+
temperature=0.1,
|
211 |
+
top_p=0.9,
|
212 |
+
)
|
213 |
+
response = outputs[0][input_ids.shape[-1]:]
|
214 |
+
raw_output = tokenizer.decode(response, skip_special_tokens=True)
|
215 |
+
return raw_output
|
216 |
+
|
217 |
+
raw_output = run_model(messages, functions)
|
218 |
+
# Check if there's a function call
|
219 |
+
function_call = postprocess_output(raw_output)
|
220 |
+
if function_call:
|
221 |
+
print(function_call)
|
222 |
+
else:
|
223 |
+
print(raw_output)
|
224 |
+
```
|
225 |
+
|
226 |
+
You should see this output, which is a function call made by the AI assistant:
|
227 |
+
```
|
228 |
+
[{'id': 'fc65a533', 'function': {'name': 'addition', 'arguments': '{"a": "4", "b": "6"}'}, 'type': 'function'}]
|
229 |
+
```
|
230 |
+
|
231 |
+
### 4. Add Executed Tool Result to Message History & Continue the Conversation
|
232 |
+
|
233 |
+
```python
|
234 |
+
if function_call:
|
235 |
+
# append the assistant tool call msg
|
236 |
+
messages.append({"role": "assistant", "tool_calls": function_call})
|
237 |
+
# append the result of the tool call in openai format, in this case, the value of add 6 to 4 is 10.
|
238 |
+
messages.append({'role': 'tool', 'tool_call_id': function_call[0]["id"], 'name': function_call[0]["function"]["name"], 'content': '10'})
|
239 |
+
raw_output = run_model(messages, functions)
|
240 |
+
# Check if there's a function call
|
241 |
+
function_call = postprocess_output(raw_output)
|
242 |
+
if function_call:
|
243 |
+
print(function_call)
|
244 |
+
else:
|
245 |
+
print(raw_output)
|
246 |
+
```
|
247 |
+
|
248 |
+
The LLM will make another call
|
249 |
+
```
|
250 |
+
[{'id': '2ffc3de4', 'function': {'name': 'addition', 'arguments': '{"a": "10", "b": "2"}'}, 'type': 'function'}]
|
251 |
```
|
252 |
|
253 |
## Training Hyperparameters
|