Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Training
|
2 |
+
|
3 |
+
This is the 10k steps English supervised-fine-tuning (SFT) model of GPT-J using SODA dataset for Chai Competition.
|
4 |
+
|
5 |
+
- **Language:** English
|
6 |
+
- **Finetuned from:** [EleutherAI / GPT-J](https://huggingface.co/EleutherAI/gpt-j-6b)
|
7 |
+
- **Code:** [Open-Assistant/model/model_training](https://github.com/LAION-AI/Open-Assistant/tree/main/model/model_training)
|
8 |
+
- **Dataset:** 10 percent from [SODA dataset](https://huggingface.co/datasets/allenai/soda)
|
9 |
+
|
10 |
+
# Why OpenAssistant framework:
|
11 |
+
- Easy to setup training with change config from dataset and model is all you need
|
12 |
+
- Data processing available for almost popular conversation datasets: SODA, Vicuna, OpenAssistant, ...
|
13 |
+
|
14 |
+
# Configuration:
|
15 |
+
|
16 |
+
You need to add this to default config file `configs/config.yaml`
|
17 |
+
|
18 |
+
data:
|
19 |
+
```
|
20 |
+
soda-only:
|
21 |
+
datasets:
|
22 |
+
- soda:
|
23 |
+
fraction: 0.1
|
24 |
+
input_max_length: 1024
|
25 |
+
```
|
26 |
+
|
27 |
+
gptj-chai:
|
28 |
+
```
|
29 |
+
dtype: fp16
|
30 |
+
log_dir: gptj-soda
|
31 |
+
model_name: EleutherAI/gpt-j-6b
|
32 |
+
output_dir: output/gptj-soda-chai
|
33 |
+
max_length: 1024
|
34 |
+
warmup_steps: 100
|
35 |
+
gradient_checkpointing: true
|
36 |
+
gradient_accumulation_steps: 1
|
37 |
+
per_device_train_batch_size: 8
|
38 |
+
per_device_eval_batch_size: 8
|
39 |
+
eval_steps: 5000
|
40 |
+
save_steps: 5000
|
41 |
+
num_train_epochs: 1
|
42 |
+
save_total_limit: 1
|
43 |
+
use_flash_attention: false
|
44 |
+
```
|
45 |
+
|
46 |
+
# Command to train:
|
47 |
+
deepspeed trainer_sft.py --local_rank=0 --configs defaults gptj-chai soda-only --cache_dir data_cache --deepspeed
|
48 |
+
|
49 |
+
# Demo code:
|
50 |
+
```python
|
51 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
52 |
+
|
53 |
+
|
54 |
+
|
55 |
+
class ChatBot():
|
56 |
+
def __init__(self, path="/mnt/hdd/duyphung/gptj-soda-chai/checkpoint-10000/"):
|
57 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
58 |
+
self.model = AutoModelForCausalLM.from_pretrained(path).half().cuda().eval()
|
59 |
+
self.model.pad_token_id = self.tokenizer.eos_token_id
|
60 |
+
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
|
61 |
+
|
62 |
+
def chat(self, message):
|
63 |
+
enc_dict = self.tokenizer(
|
64 |
+
message,
|
65 |
+
return_tensors='pt'
|
66 |
+
)
|
67 |
+
for x in enc_dict:
|
68 |
+
enc_dict[x] = enc_dict[x].cuda()
|
69 |
+
chat_history_ids = self.model.generate(
|
70 |
+
input_ids=enc_dict['input_ids'],
|
71 |
+
attention_mask=enc_dict['attention_mask'],
|
72 |
+
max_new_tokens=64,
|
73 |
+
temperature=0.7,
|
74 |
+
do_sample=True,
|
75 |
+
top_k=0,
|
76 |
+
top_p=0.95,
|
77 |
+
)
|
78 |
+
out = chat_history_ids[:, enc_dict['input_ids'].shape[-1]:][0]
|
79 |
+
return self.tokenizer.decode(out, skip_special_tokens=True)
|
80 |
+
|
81 |
+
|
82 |
+
if __name__ == "__main__":
|
83 |
+
bot_name = 'Bot:'
|
84 |
+
prompt = "<|prompter|>"
|
85 |
+
chat_history = []
|
86 |
+
|
87 |
+
bot = ChatBot()
|
88 |
+
while True:
|
89 |
+
message = input("Me: ")
|
90 |
+
chat_history.append(f'Me: {message}')
|
91 |
+
prompt = prompt + message + "<|endoftext|><|assistant|>"
|
92 |
+
response = bot.chat(prompt)
|
93 |
+
print(f'{bot_name} {response}')
|
94 |
+
prompt = prompt + response + "<|endoftext|><|prompter|>"
|
95 |
+
```
|
96 |
+
|