Introduction

Our Model is fine-tuned on Llama-2 7B model on Text-2-SQL Dataset based on Alpaca format described by Stanford. We have used QLora, Bits&Bytes, Accelerate and Transformers Library to implement PEFT concept. For more information, please visit : https://github.com/akshayhedaoo1/Llama-2-7b-chat-finetune-for-text2sql/tree/Data-Science

Inference

!pip install transformers accelerate xformers bitsandbytes

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

tokenizer = AutoTokenizer.from_pretrained("ekshat/Llama-2-7b-chat-finetune-for-text2sql")

# Loading model in 4 bit precision
model = AutoModelForCausalLM.from_pretrained("ekshat/Llama-2-7b-chat-finetune-for-text2sql", load_in_4bit=True)

context = "CREATE TABLE head (name VARCHAR, born_state VARCHAR, age VARCHAR)"
question = "List the name, born state and age of the heads of departments ordered by age."

prompt = f"""Below is an context that describes a sql query, paired with an question that provides further information. Write an answer that appropriately completes the request.
### Context:
{context}
### Question:
{question}
### Answer:"""

pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, max_length=200)
result = pipe(prompt)
print(result[0]['generated_text'])

Model Information

• model_name = "NousResearch/Llama-2-7b-chat-hf"

• dataset_name = "ekshat/text-2-sql-with-context"

QLoRA parameters

• lora_r = 64

• lora_alpha = 16

• lora_dropout = 0.1

BitsAndBytes parameters

• use_4bit = True

• bnb_4bit_compute_dtype = "float16"

• bnb_4bit_quant_type = "nf4"

• use_nested_quant = False

Training Arguments parameters

• num_train_epochs = 1

• fp16 = False

• bf16 = False

• per_device_train_batch_size = 8

• per_device_eval_batch_size = 4

• gradient_accumulation_steps = 1

• gradient_checkpointing = True

• max_grad_norm = 0.3

• learning_rate = 2e-4

• weight_decay = 0.001

• optim = "paged_adamw_32bit"

• lr_scheduler_type = "cosine"

• max_steps = -1

• warmup_ratio = 0.03

• group_by_length = True

• save_steps = 0

• logging_steps = 25

SFT parameters

• max_seq_length = None

• packing = False