HolyYi-9B / README.md
nbeerbower's picture
Update README.md
b4fc76f verified
---
library_name: transformers
license: apache-2.0
datasets:
- nbeerbower/bible-dpo
base_model:
- 01-ai/Yi-1.5-9B-Chat
---
# HolyYi-9B
[01-ai/Yi-1.5-9B-Chat](https://huggingface.co/01-ai/Yi-1.5-9B-Chat) finetuned on [nbeerbower/bible-dpo](https://huggingface.co/datasets/nbeerbower/bible-dpo).
### Method
Finetuned using an A100 on Google Colab.
[Fine-Tune Your Own Llama 2 Model in a Colab Notebook](https://mlabonne.github.io/blog/posts/Fine_Tune_Your_Own_Llama_2_Model_in_a_Colab_Notebook.html)
### Configuration
Dataset preparation, system prompt:
```python
def chatml_format(example):
# Format system
system = ""
systemMessage = "Recite the given verse from the Bible."
system = "<|im_start|>system\n" + systemMessage + "<|im_end|>\n"
# Format instruction
prompt = "<|im_start|>user\nRecite " + example['citation'] + "<|im_end|>\n<|im_start|>assistant\n"
# Format chosen answer
chosen = example['text'] + "<|im_end|>\n"
# Format rejected answer
rejected = example['rejected'] + "<|im_end|>\n"
return {
"prompt": system + prompt,
"chosen": chosen,
"rejected": rejected,
}
dataset = load_dataset("nbeerbower/bible-dpo")['train']
# Save columns
original_columns = dataset.column_names
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
# Format dataset
dataset = dataset.map(
chatml_format,
remove_columns=original_columns
)
```
LoRA, model, and training settings:
```python
# LoRA configuration
peft_config = LoraConfig(
r=16,
lora_alpha=16,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)
# Model to fine-tune
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
load_in_4bit=True
)
model.config.use_cache = False
# Reference model
ref_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
load_in_4bit=True
)
# Training arguments
training_args = TrainingArguments(
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
gradient_checkpointing=True,
learning_rate=5e-5,
lr_scheduler_type="cosine",
max_steps=5000,
save_strategy="no",
logging_steps=1,
output_dir=new_model,
optim="paged_adamw_32bit",
warmup_steps=100,
bf16=True,
report_to="wandb",
)
# Create DPO trainer
dpo_trainer = DPOTrainer(
model,
ref_model,
args=training_args,
train_dataset=dataset,
tokenizer=tokenizer,
peft_config=peft_config,
beta=0.1,
max_prompt_length=512,
max_length=1536,
force_use_ref_model=True
)
```