Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# How to use
|
2 |
+
|
3 |
+
We write our prompts in the ChatML format.
|
4 |
+
|
5 |
+
### With vLLM (recommended for much faster inference)
|
6 |
+
|
7 |
+
<details><summary>Install vLLM</summary>
|
8 |
+
```bash
|
9 |
+
pip install vllm
|
10 |
+
```
|
11 |
+
[Reference](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
|
12 |
+
</details>
|
13 |
+
|
14 |
+
```python
|
15 |
+
from vllm import LLM, SamplingParams
|
16 |
+
model_name = "lightblue/jod"
|
17 |
+
llm = LLM(model=model_name)
|
18 |
+
|
19 |
+
SYSTEM_MESSAGE = "You are a helpful assistant."
|
20 |
+
def process_chat_history(next_user_msg, text_chat_history = []):
|
21 |
+
prompt_text = "<|im_start|>system\n"
|
22 |
+
prompt_text += SYSTEM_MESSAGE
|
23 |
+
prompt_text += "<|im_end|>\n\n"
|
24 |
+
|
25 |
+
for user_msg, ai_msg in text_chat_history:
|
26 |
+
prompt_text += "<|im_start|>user\n"
|
27 |
+
prompt_text += user_msg
|
28 |
+
prompt_text += "<|im_end|>\n\n"
|
29 |
+
prompt_text += "<|im_start|>assistant\n"
|
30 |
+
prompt_text += ai_msg
|
31 |
+
prompt_text += "<|im_end|>\n\n"
|
32 |
+
|
33 |
+
prompt_text += "<|im_start|>user\n"
|
34 |
+
prompt_text += next_user_msg
|
35 |
+
prompt_text += "<|im_end|>\n\n"
|
36 |
+
prompt_text += "<|im_start|>assistant\n"
|
37 |
+
return prompt_text
|
38 |
+
|
39 |
+
user_prompt = "鏃ユ湰銇竴鐣珮銇勫北銇紵"
|
40 |
+
prompt = process_chat_history(user_prompt)
|
41 |
+
sampling_params = SamplingParams(temperature=0, max_tokens=528)
|
42 |
+
outputs = llm.generate(prompt, sampling_params)
|
43 |
+
bot_message = outputs[0].outputs[0].text.strip()
|
44 |
+
print(bot_message)
|
45 |
+
```
|
46 |
+
|
47 |
+
|
48 |
+
### With Huggingface
|
49 |
+
|
50 |
+
```python
|
51 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
52 |
+
|
53 |
+
model_name = "lightblue/jod"
|
54 |
+
|
55 |
+
tokenizer = AutoTokenizer.from_pretrained(model_dir)
|
56 |
+
model = AutoModelForCausalLM.from_pretrained(
|
57 |
+
model_dir, torch_dtype=torch.bfloat16, device_map='auto', load_in_4bit=True,
|
58 |
+
)
|
59 |
+
|
60 |
+
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
61 |
+
|
62 |
+
SYSTEM_MESSAGE = "You are a helpful assistant."
|
63 |
+
def process_chat_history(next_user_msg, text_chat_history = []):
|
64 |
+
prompt_text = "<|im_start|>system\n"
|
65 |
+
prompt_text += SYSTEM_MESSAGE
|
66 |
+
prompt_text += "<|im_end|>\n\n"
|
67 |
+
|
68 |
+
for user_msg, ai_msg in text_chat_history:
|
69 |
+
prompt_text += "<|im_start|>user\n"
|
70 |
+
prompt_text += user_msg
|
71 |
+
prompt_text += "<|im_end|>\n\n"
|
72 |
+
prompt_text += "<|im_start|>assistant\n"
|
73 |
+
prompt_text += ai_msg
|
74 |
+
prompt_text += "<|im_end|>\n\n"
|
75 |
+
|
76 |
+
prompt_text += "<|im_start|>user\n"
|
77 |
+
prompt_text += next_user_msg
|
78 |
+
prompt_text += "<|im_end|>\n\n"
|
79 |
+
prompt_text += "<|im_start|>assistant\n"
|
80 |
+
return prompt_text
|
81 |
+
|
82 |
+
user_prompt = "鏃ユ湰銇竴鐣珮銇勫北銇紵"
|
83 |
+
prompt = process_chat_history(user_prompt)
|
84 |
+
bot_message = pipe(do_closed_qa(test_article, question), max_new_tokens=128, temperature=0)[0]["generated_text"]
|
85 |
+
print(bot_message)
|
86 |
+
```
|
87 |
+
|
88 |
+
|
89 |
+
# Training datasets
|
90 |
+
This model was trained using the ChatML format, so it should be used for inference using the ChatML chatbot format.
|
91 |
+
We chose this format as the base model ([Open-Orca/Mistral-7B-SlimOrca](https://huggingface.co/Open-Orca/Mistral-7B-SlimOrca)) was trained with this format, and we find the chatbot format more compelling for practical use compared to the Alpaca style instruction format.
|
92 |
+
|
93 |
+
* [JASTER](https://github.com/llm-jp/llm-jp-eval)
|
94 |
+
* [kunishou/oasst1-89k-ja](https://huggingface.co/datasets/kunishou/oasst1-89k-ja/)
|
95 |
+
* [kunishou/databricks-dolly-15k-ja](https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja/)
|
96 |
+
|
97 |
+
We trained for 1 epoch using the following Axolotl config. (Early stopping was not performed during our training.)
|
98 |
+
<details><summary>Axolotl config .yaml</summary>
|
99 |
+
|
100 |
+
```yaml
|
101 |
+
base_model: Open-Orca/Mistral-7B-SlimOrca
|
102 |
+
base_model_config: Open-Orca/Mistral-7B-SlimOrca
|
103 |
+
model_type: MistralForCausalLM
|
104 |
+
tokenizer_type: LlamaTokenizer
|
105 |
+
is_mistral_derived_model: true
|
106 |
+
|
107 |
+
load_in_8bit: false
|
108 |
+
load_in_4bit: true
|
109 |
+
strict: false
|
110 |
+
|
111 |
+
datasets:
|
112 |
+
- path: ./data/jaster_plus.jsonl
|
113 |
+
ds_type: json # see other options below
|
114 |
+
type: sharegpt
|
115 |
+
conversation: chatml
|
116 |
+
dataset_prepared_path: false
|
117 |
+
val_set_size: 0.002
|
118 |
+
output_dir: ./train_output/openorca-mistral-jaster-1epoch
|
119 |
+
|
120 |
+
use_wandb: true
|
121 |
+
wandb_project: \<HIDDEN\>
|
122 |
+
wandb_entity: \<HIDDEN\>
|
123 |
+
|
124 |
+
debug:
|
125 |
+
|
126 |
+
adapter: qlora
|
127 |
+
lora_model_dir:
|
128 |
+
|
129 |
+
sequence_len: 4096
|
130 |
+
sample_packing: true
|
131 |
+
pad_to_sequence_len: true
|
132 |
+
|
133 |
+
lora_r: 32
|
134 |
+
lora_alpha: 16
|
135 |
+
lora_dropout: 0.05
|
136 |
+
lora_target_linear: true
|
137 |
+
lora_fan_in_fan_out:
|
138 |
+
lora_target_modules:
|
139 |
+
- gate_proj
|
140 |
+
- down_proj
|
141 |
+
- up_proj
|
142 |
+
- q_proj
|
143 |
+
- v_proj
|
144 |
+
- k_proj
|
145 |
+
- o_proj
|
146 |
+
|
147 |
+
gradient_accumulation_steps: 1
|
148 |
+
micro_batch_size: 10
|
149 |
+
eval_batch_size: 4
|
150 |
+
num_epochs: 1
|
151 |
+
optimizer: adamw_bnb_8bit
|
152 |
+
lr_scheduler: cosine
|
153 |
+
learning_rate: 0.0002
|
154 |
+
|
155 |
+
train_on_inputs: false
|
156 |
+
group_by_length: false
|
157 |
+
bf16: true
|
158 |
+
fp16: false
|
159 |
+
tf32: false
|
160 |
+
|
161 |
+
gradient_checkpointing: true
|
162 |
+
early_stopping_patience: 10
|
163 |
+
resume_from_checkpoint:
|
164 |
+
local_rank:
|
165 |
+
logging_steps: 1
|
166 |
+
xformers_attention:
|
167 |
+
flash_attention: true
|
168 |
+
|
169 |
+
warmup_steps: 10
|
170 |
+
eval_steps: 10
|
171 |
+
eval_table_size: 5
|
172 |
+
eval_table_max_new_tokens: 128
|
173 |
+
save_steps: 10
|
174 |
+
debug:
|
175 |
+
deepspeed:
|
176 |
+
weight_decay: 0.0
|
177 |
+
fsdp:
|
178 |
+
fsdp_config:
|
179 |
+
special_tokens:
|
180 |
+
bos_token: "<s>"
|
181 |
+
eos_token: "</s>"
|
182 |
+
unk_token: "<unk>"
|
183 |
+
```
|
184 |
+
|
185 |
+
</details>
|