Update README.md
Browse files
README.md
CHANGED
@@ -4,4 +4,64 @@ license: apache-2.0
|
|
4 |
|
5 |
Modelo entrenado con DPO
|
6 |
|
7 |
-
Merge de dos modelos
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
Modelo entrenado con DPO
|
6 |
|
7 |
+
Merge de dos modelos
|
8 |
+
|
9 |
+
codigo de train:
|
10 |
+
|
11 |
+
# LoRA configuration
|
12 |
+
peft_config = LoraConfig(
|
13 |
+
r=16,
|
14 |
+
lora_alpha=16,
|
15 |
+
lora_dropout=0.05,
|
16 |
+
bias="none",
|
17 |
+
task_type="CAUSAL_LM",
|
18 |
+
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
|
19 |
+
)
|
20 |
+
|
21 |
+
# Model to fine-tune
|
22 |
+
model = AutoModelForCausalLM.from_pretrained(
|
23 |
+
model_name,
|
24 |
+
torch_dtype=torch.float16,
|
25 |
+
load_in_4bit=True
|
26 |
+
)
|
27 |
+
model.config.use_cache = False
|
28 |
+
|
29 |
+
# Reference model
|
30 |
+
ref_model = AutoModelForCausalLM.from_pretrained(
|
31 |
+
model_name,
|
32 |
+
torch_dtype=torch.float16,
|
33 |
+
load_in_4bit=True
|
34 |
+
)
|
35 |
+
|
36 |
+
# Training arguments
|
37 |
+
training_args = TrainingArguments(
|
38 |
+
per_device_train_batch_size=4,
|
39 |
+
gradient_accumulation_steps=4,
|
40 |
+
gradient_checkpointing=True,
|
41 |
+
learning_rate=5e-5,
|
42 |
+
lr_scheduler_type="cosine",
|
43 |
+
max_steps=200,
|
44 |
+
save_strategy="no",
|
45 |
+
logging_steps=1,
|
46 |
+
output_dir=new_model,
|
47 |
+
optim="paged_adamw_32bit",
|
48 |
+
warmup_steps=100,
|
49 |
+
bf16=True,
|
50 |
+
report_to="wandb",
|
51 |
+
)
|
52 |
+
|
53 |
+
# Create DPO trainer
|
54 |
+
dpo_trainer = DPOTrainer(
|
55 |
+
model,
|
56 |
+
ref_model,
|
57 |
+
args=training_args,
|
58 |
+
train_dataset=dataset,
|
59 |
+
tokenizer=tokenizer,
|
60 |
+
peft_config=peft_config,
|
61 |
+
beta=0.1,
|
62 |
+
max_prompt_length=1024,
|
63 |
+
max_length=1536,
|
64 |
+
)
|
65 |
+
|
66 |
+
# Fine-tune model with DPO
|
67 |
+
dpo_trainer.train()
|