Upload train_human.py
Browse files- train_human.py +146 -0
train_human.py
ADDED
|
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# %%
|
| 2 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 3 |
+
|
| 4 |
+
from datasets import load_dataset, Dataset
|
| 5 |
+
|
| 6 |
+
from trl import DPOTrainer, DPOConfig
|
| 7 |
+
from peft import LoraConfig
|
| 8 |
+
from peft import prepare_model_for_kbit_training
|
| 9 |
+
import torch
|
| 10 |
+
|
| 11 |
+
import pandas as pd
|
| 12 |
+
|
| 13 |
+
# %%
|
| 14 |
+
dataset = load_dataset("Undi95/Weyaxi-humanish-dpo-project-noemoji")["train"]
|
| 15 |
+
|
| 16 |
+
model_name = "Undi95/Meta-Llama-3.1-8B-Claude-bf16"
|
| 17 |
+
|
| 18 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 19 |
+
tokenizer.padding_side = "right"
|
| 20 |
+
|
| 21 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 22 |
+
|
| 23 |
+
# %%
|
| 24 |
+
tokenizer.chat_template = "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}"
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
# %%
|
| 28 |
+
dataset2 = load_dataset("ResplendentAI/NSFW_RP_Format_DPO")['train']
|
| 29 |
+
|
| 30 |
+
# %%
|
| 31 |
+
dataset = dataset.to_pandas(
|
| 32 |
+
)
|
| 33 |
+
dataset2 = dataset2.to_pandas()
|
| 34 |
+
|
| 35 |
+
dataset = Dataset.from_pandas(pd.concat([dataset.sample(400), dataset2]).sample(frac=1))
|
| 36 |
+
|
| 37 |
+
# %%
|
| 38 |
+
def template_prompt(system, prompt):
|
| 39 |
+
if system is None:
|
| 40 |
+
messages = [
|
| 41 |
+
{"role": "user", "content": prompt},
|
| 42 |
+
]
|
| 43 |
+
else:
|
| 44 |
+
messages = [
|
| 45 |
+
{
|
| 46 |
+
"role": "system",
|
| 47 |
+
"content": system,
|
| 48 |
+
},
|
| 49 |
+
{"role": "user", "content": prompt},
|
| 50 |
+
]
|
| 51 |
+
prompt = tokenizer.apply_chat_template(
|
| 52 |
+
messages, tokenize=False, add_generation_prompt=False
|
| 53 |
+
)
|
| 54 |
+
return prompt
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def template_answer(answer):
|
| 58 |
+
messages = [
|
| 59 |
+
{
|
| 60 |
+
"role": "assistant",
|
| 61 |
+
"content": answer,
|
| 62 |
+
},
|
| 63 |
+
]
|
| 64 |
+
answer = tokenizer.apply_chat_template(
|
| 65 |
+
messages, tokenize=False, add_generation_prompt=False
|
| 66 |
+
)
|
| 67 |
+
return answer
|
| 68 |
+
|
| 69 |
+
# %%
|
| 70 |
+
# create new columns
|
| 71 |
+
dataset = dataset.map(
|
| 72 |
+
lambda x: {
|
| 73 |
+
"prompt": template_prompt(None, x["prompt"]).replace("<|start_header_id|>assistant<|end_header_id|>\n\n", "")
|
| 74 |
+
}, # change this according to the dataset!!!
|
| 75 |
+
)
|
| 76 |
+
|
| 77 |
+
# %%
|
| 78 |
+
dataset = dataset.map(
|
| 79 |
+
lambda x: {"chosen": template_answer(x["chosen"]).replace('<|begin_of_text|>', '').replace('><|start_header_id|>assistant<|end_header_id|>\n\n', '>')},
|
| 80 |
+
)
|
| 81 |
+
dataset = dataset.map(
|
| 82 |
+
lambda x: {"rejected": template_answer(x["rejected"]).replace('<|begin_of_text|>', '').replace('><|start_header_id|>assistant<|end_header_id|>\n\n', '>')},
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
# %%
|
| 86 |
+
dataset[0]
|
| 87 |
+
|
| 88 |
+
# %%
|
| 89 |
+
# LoRA configuration
|
| 90 |
+
peft_config = LoraConfig(
|
| 91 |
+
r=16,
|
| 92 |
+
lora_alpha=32,
|
| 93 |
+
lora_dropout=0.05,
|
| 94 |
+
bias="none",
|
| 95 |
+
task_type="CAUSAL_LM",
|
| 96 |
+
target_modules=[
|
| 97 |
+
"k_proj",
|
| 98 |
+
"gate_proj",
|
| 99 |
+
"v_proj",
|
| 100 |
+
"up_proj",
|
| 101 |
+
"q_proj",
|
| 102 |
+
"o_proj",
|
| 103 |
+
"down_proj",
|
| 104 |
+
],
|
| 105 |
+
)
|
| 106 |
+
|
| 107 |
+
# Model to fine-tune
|
| 108 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 109 |
+
model_name,
|
| 110 |
+
torch_dtype=torch.float16,
|
| 111 |
+
load_in_4bit=True,
|
| 112 |
+
device_map="auto",
|
| 113 |
+
)
|
| 114 |
+
model.config.use_cache = False
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
model.gradient_checkpointing_enable()
|
| 118 |
+
model = prepare_model_for_kbit_training(model)
|
| 119 |
+
|
| 120 |
+
# %%
|
| 121 |
+
output_name = f"checkpoints/exp_human_{model_name}"
|
| 122 |
+
|
| 123 |
+
training_args = DPOConfig(
|
| 124 |
+
per_device_train_batch_size=1,
|
| 125 |
+
gradient_accumulation_steps=4,
|
| 126 |
+
num_train_epochs=1,
|
| 127 |
+
gradient_checkpointing=True,
|
| 128 |
+
output_dir=output_name,
|
| 129 |
+
logging_steps=1,
|
| 130 |
+
max_steps=50
|
| 131 |
+
)
|
| 132 |
+
|
| 133 |
+
trainer = DPOTrainer(
|
| 134 |
+
model,
|
| 135 |
+
ref_model=None,
|
| 136 |
+
train_dataset=dataset,
|
| 137 |
+
tokenizer=tokenizer,
|
| 138 |
+
args=training_args,
|
| 139 |
+
peft_config=peft_config,
|
| 140 |
+
)
|
| 141 |
+
|
| 142 |
+
trainer.train()
|
| 143 |
+
|
| 144 |
+
trainer.save_model(output_name)
|
| 145 |
+
|
| 146 |
+
|