llama 3 experiments
Collection
Merge and finetune experiments with llama 3, aiming for decensorship and benchmark performance.
•
17 items
•
Updated
This model is based on Llama-3-8b, and is governed by META LLAMA 3 COMMUNITY LICENSE AGREEMENT
nbeerbower/llama-3-bophades-v3-8B finetuned on tasksource/ScienceQA_text_only.
Finetuned using an A100 on Google Colab.
Fine-Tune Your Own Llama 2 Model in a Colab Notebook
Dataset preparation, system prompt:
def get_correct_answer(example):
answerIdx = example['answer']
choices = example['choices']
return choices[answerIdx]
def get_wrong_answer(example):
choices = example['choices']
answerIdx = example['answer']
for i in range(len(choices)):
if i != answerIdx:
return choices[i]
def chatml_format(example):
# Format system
systemMessage = "Read the following lecture, then answer the question."
system = "<|im_start|>system\n" + systemMessage + "<|im_end|>\n"
# Format instruction
instruction = ""
if example.get('lecture'):
instruction = "Lecture: " + example['lecture'] + "\nQuestion: "
else:
instruction = "Question: "
instruction += example['question']
# Format prompt
prompt = "<|im_start|>user\n" + instruction + "<|im_end|>\n<|im_start|>assistant\n"
# Format chosen answer
chosen = get_correct_answer(example) + "<|im_end|>\n"
# Format rejected answer
rejected = get_wrong_answer(example) + "<|im_end|>\n"
return {
"prompt": system + prompt,
"chosen": chosen,
"rejected": rejected,
}
dataset = load_dataset("tasksource/ScienceQA_text_only")['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:
# 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=1000,
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=1024,
max_length=1536,
force_use_ref_model=True
)
Base model
nbeerbower/llama-3-wissenschaft-8B