Paul Rock
commited on
Commit
•
d188338
1
Parent(s):
99bdfc7
Readme updated
Browse files- .gitignore +1 -0
- README.md +191 -1
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
/.idea/
|
README.md
CHANGED
@@ -14,4 +14,194 @@ language:
|
|
14 |
pipeline_tag: conversational
|
15 |
tags:
|
16 |
- ruGPT
|
17 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
pipeline_tag: conversational
|
15 |
tags:
|
16 |
- ruGPT
|
17 |
+
---
|
18 |
+
|
19 |
+
# ruGPT-3.5 13B GGML
|
20 |
+
|
21 |
+
Welcome to the adapter-only version of ruGPT-3.5 13B GGML. This model is built upon the foundation of [ruGPT-3.5-13B](https://huggingface.co/ai-forever/ruGPT-3.5-13B).
|
22 |
+
|
23 |
+
📌 Important: This model was trained using settings identical to [GigaSaiga](https://huggingface.co/IlyaGusev/gigasaiga_lora), but incorporates additional dataset.
|
24 |
+
|
25 |
+
🔗 Training code is [here](https://github.com/EvilFreelancer/ruGPT-3.5-13B-lora).
|
26 |
+
|
27 |
+
## Code sample
|
28 |
+
|
29 |
+
```python
|
30 |
+
from llm_rs import AutoModel, GenerationConfig as GConfig
|
31 |
+
from transformers import AutoTokenizer, GenerationConfig
|
32 |
+
|
33 |
+
MODEL_NAME = "evilfreelancer/ruGPT-3.5-13B-ggml"
|
34 |
+
DEFAULT_MESSAGE_TEMPLATE = "<s>{role}\n{content}</s>\n"
|
35 |
+
DEFAULT_SYSTEM_PROMPT = "Ты — ruGPT-3.5, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им."
|
36 |
+
|
37 |
+
class Conversation:
|
38 |
+
def __init__(
|
39 |
+
self,
|
40 |
+
message_template=DEFAULT_MESSAGE_TEMPLATE,
|
41 |
+
system_prompt=DEFAULT_SYSTEM_PROMPT,
|
42 |
+
start_token_id=2,
|
43 |
+
bot_token_id=46787
|
44 |
+
):
|
45 |
+
self.message_template = message_template
|
46 |
+
self.start_token_id = start_token_id
|
47 |
+
self.bot_token_id = bot_token_id
|
48 |
+
self.messages = [{
|
49 |
+
"role": "system",
|
50 |
+
"content": system_prompt
|
51 |
+
}]
|
52 |
+
|
53 |
+
def get_start_token_id(self):
|
54 |
+
return self.start_token_id
|
55 |
+
|
56 |
+
def get_bot_token_id(self):
|
57 |
+
return self.bot_token_id
|
58 |
+
|
59 |
+
def add_user_message(self, message):
|
60 |
+
self.messages.append({
|
61 |
+
"role": "user",
|
62 |
+
"content": message
|
63 |
+
})
|
64 |
+
|
65 |
+
def add_bot_message(self, message):
|
66 |
+
self.messages.append({
|
67 |
+
"role": "bot",
|
68 |
+
"content": message
|
69 |
+
})
|
70 |
+
|
71 |
+
def get_prompt(self, tokenizer):
|
72 |
+
final_text = ""
|
73 |
+
for message in self.messages:
|
74 |
+
message_text = self.message_template.format(**message)
|
75 |
+
final_text += message_text
|
76 |
+
final_text += tokenizer.decode([self.start_token_id, self.bot_token_id])
|
77 |
+
return final_text.strip()
|
78 |
+
|
79 |
+
|
80 |
+
def generate(model, tokenizer, prompt, generation_config):
|
81 |
+
data = tokenizer(prompt, return_tensors="pt")
|
82 |
+
output = model.generate(
|
83 |
+
prompt=prompt,
|
84 |
+
generation_config=generation_config
|
85 |
+
).text
|
86 |
+
# print("output", output)
|
87 |
+
output_ids = tokenizer(output, return_tensors="pt")['input_ids'][0]
|
88 |
+
# print("output_ids", output_ids)
|
89 |
+
# output_ids = output_ids[len(data["input_ids"][0]):]
|
90 |
+
# print("output_ids", output_ids)
|
91 |
+
output = tokenizer.decode(output_ids, skip_special_tokens=True)
|
92 |
+
# print("output_ids", output)
|
93 |
+
return output.strip()
|
94 |
+
|
95 |
+
# Load base model
|
96 |
+
model = AutoModel.from_pretrained(
|
97 |
+
MODEL_NAME,
|
98 |
+
model_file="ruGPT-3.5-13B-lora-q4_0.bin",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Init basic tokenizer
|
102 |
+
tokenizer = AutoTokenizer.from_pretrained('ai-forever/ruGPT-3.5-13B', use_fast=False)
|
103 |
+
generation_config = GenerationConfig.from_pretrained('ai-forever/ruGPT-3.5-13B')
|
104 |
+
|
105 |
+
# Instantiate generator config
|
106 |
+
g_config = GConfig()
|
107 |
+
g_config.top_p = generation_config['top_p']
|
108 |
+
g_config.top_k = generation_config['top_k']
|
109 |
+
g_config.repetition_penalty = generation_config['repetition_penalty']
|
110 |
+
g_config.temperature = generation_config['temperature']
|
111 |
+
g_config.max_new_tokens = 200
|
112 |
+
|
113 |
+
# Start conversation
|
114 |
+
conversation = Conversation()
|
115 |
+
while True:
|
116 |
+
user_message = input("User: ")
|
117 |
+
|
118 |
+
if user_message.strip() == "/reset":
|
119 |
+
conversation = Conversation()
|
120 |
+
print("History reset completed!")
|
121 |
+
continue
|
122 |
+
|
123 |
+
if user_message.strip() == "":
|
124 |
+
continue
|
125 |
+
|
126 |
+
conversation.add_user_message(user_message)
|
127 |
+
prompt = conversation.get_prompt(tokenizer)
|
128 |
+
output = generate(
|
129 |
+
model=model,
|
130 |
+
tokenizer=tokenizer,
|
131 |
+
prompt=prompt,
|
132 |
+
generation_config=g_config
|
133 |
+
)
|
134 |
+
conversation.add_bot_message(output)
|
135 |
+
print("ruGPT-3.5:", output)
|
136 |
+
print()
|
137 |
+
print("==============================")
|
138 |
+
print()
|
139 |
+
```
|
140 |
+
|
141 |
+
## Examples
|
142 |
+
|
143 |
+
```text
|
144 |
+
User: Привет! Почему трава зелёная?
|
145 |
+
ruGPT-3.5: Трава зелёная благодаря хлорофиллу - пигменту растений, который поглощает световые волны определенной длины и преобразует их в энергию химических связей. Хлорофилл также обеспечивает растения питательными веществами из почвы через фотосинтез.
|
146 |
+
```
|
147 |
+
|
148 |
+
```text
|
149 |
+
User: Придумай стишок про нейросети.
|
150 |
+
ruGPT-3.5: Нейросеть – это чудо науки,
|
151 |
+
Она способна решать любые задачи.
|
152 |
+
И если ты хочешь стать программистом,
|
153 |
+
То лучше всего изучать нейронные сети.
|
154 |
+
```
|
155 |
+
|
156 |
+
```text
|
157 |
+
User: Высота само�� высокой горы на Земле?
|
158 |
+
ruGPT-3.5: Самая высокая гора на Земле находится в Гималаях и называется Эверест (8848 метров).
|
159 |
+
```
|
160 |
+
|
161 |
+
```text
|
162 |
+
User: Сколько максимум лет может прожить галопогосская черепаха?
|
163 |
+
ruGPT-3.5: Галопогосские черепахи могут жить до 100 лет и более. Они являются одними из самых долгоживущих животных на планете.
|
164 |
+
```
|
165 |
+
|
166 |
+
## 📚 Training Datasets
|
167 |
+
|
168 |
+
The datasets utilized for training this model are consistent with those used for [Saiga-2](https://github.com/IlyaGusev/rulm).
|
169 |
+
|
170 |
+
Here's the comprehensive list:
|
171 |
+
|
172 |
+
- [ru_turbo_alpaca](https://huggingface.co/datasets/IlyaGusev/ru_turbo_alpaca)
|
173 |
+
- [ru_turbo_alpaca_evol_instruct](https://huggingface.co/datasets/IlyaGusev/ru_turbo_alpaca_evol_instruct)
|
174 |
+
- [ru_turbo_saiga](https://huggingface.co/datasets/IlyaGusev/ru_turbo_saiga)
|
175 |
+
- [ru_sharegpt_cleaned](https://huggingface.co/datasets/IlyaGusev/ru_sharegpt_cleaned)
|
176 |
+
- [oasst1_ru_main_branch](https://huggingface.co/datasets/IlyaGusev/oasst1_ru_main_branch)
|
177 |
+
- [gpt_roleplay_realm](https://huggingface.co/datasets/IlyaGusev/gpt_roleplay_realm)
|
178 |
+
- [ru_instruct_gpt4](https://huggingface.co/datasets/lksy/ru_instruct_gpt4)
|
179 |
+
|
180 |
+
## 🛠 Training Procedure
|
181 |
+
|
182 |
+
The following `bitsandbytes` quantization config was used during training:
|
183 |
+
|
184 |
+
- quant_method: bitsandbytes
|
185 |
+
- load_in_8bit: True
|
186 |
+
- load_in_4bit: False
|
187 |
+
- llm_int8_threshold: 6.0
|
188 |
+
- llm_int8_skip_modules: None
|
189 |
+
- llm_int8_enable_fp32_cpu_offload: False
|
190 |
+
- llm_int8_has_fp16_weight: False
|
191 |
+
- bnb_4bit_quant_type: fp4
|
192 |
+
- bnb_4bit_use_double_quant: False
|
193 |
+
- bnb_4bit_compute_dtype: float32
|
194 |
+
|
195 |
+
## ⚙️ Framework Versions
|
196 |
+
|
197 |
+
Ensure you have the following framework versions for compatibility:
|
198 |
+
|
199 |
+
- PyTorch 2.1.0
|
200 |
+
- PEFT 0.5.0
|
201 |
+
- bitsandbytes 0.41.1
|
202 |
+
- transformers 4.34.0
|
203 |
+
|
204 |
+
## Links
|
205 |
+
|
206 |
+
- https://t.me/evilfreelancer
|
207 |
+
- https://dzen.ru/evilfreelancer
|