phi3-mini-instruct-alpaca-model

Uploaded model

• Developed by: YuvrajSingh9886
• License: apache-2.0
• Finetuned from model : unsloth/phi-3-mini-4k-instruct-bnb-4bit

<!-- Provide a quick summary of what the model is/does. -->

It's a fine-tuned version of Phi-2 model by Microsoft on Alpaca-Cleaned-52k.

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

The above model, with applicable changes to the generation_config file, passed to model.generate() function can lead to the generation of better results which could then be used for Health Counseling Chatbot dev.

Bias, Risks, and Limitations

The model was developed as a proof-of-concept type hobby project and is not intended to be used without careful consideration of its implications.

[More Information Needed]

How to Get Started with the Model

Use the code below to get started with the model.

Load in the model using the BitsandBytes library

pip install bitsandbytes

Load model from Hugging Face Hub with model name and bitsandbytes configuration

def load_model_tokenizer(model_name: str, bnb_config: BitsAndBytesConfig) -> Tuple[AutoModelForCausalLM, AutoTokenizer]:
    """
    Load the model and tokenizer from the HuggingFace model hub using quantization.

    Args:
        model_name (str): The name of the model.
        bnb_config (BitsAndBytesConfig): The quantization configuration of BitsAndBytes.

    Returns:
        Tuple[AutoModelForCausalLM, AutoTokenizer]: The model and tokenizer.
    """


    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        quantization_config = bnb_config,
        # device_map = "auto",
        torch_dtype="auto",
        trust_remote_code=True
    )

    tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token = True, trust_remote_code=True)

    tokenizer.pad_token = tokenizer.eos_token

    return model, tokenizer


bnb_config = BitsAndBytesConfig(
        load_in_4bit = load_in_4bit,
        bnb_4bit_use_double_quant = bnb_4bit_use_double_quant,
        bnb_4bit_quant_type = bnb_4bit_quant_type,
        bnb_4bit_compute_dtype = bnb_4bit_compute_dtype,
    )

model, tokenizer = load_model_tokenizer(model_name, bnb_config)

new_model = "YuvrajSingh9886/medicinal-QnA-phi2-custom"

prompt = "I have been feeling more and more down for over a month. I have started having trouble sleeping due to panic attacks, but they are almost never triggered by something that I know of."

tokens = tokenizer(f"### Question: {prompt}", return_tensors='pt').to('cuda')
tokenizer.pad_token = tokenizer.eos_token
outputs = model.generate(**tokens, max_new_tokens=1024, num_beams=5,
    no_repeat_ngram_size=2,
    early_stopping=True
                         )
print(tokenizer.batch_decode(outputs,skip_special_tokens=True)[0])

Training Details

Training Data

Hardware

Epcohs: 10 Hardware: (1) RTX 4090 (24GB VRAM) 48GB 8vCPU (RAM) Hard Disk: 40GB

[More Information Needed]

Training Procedure

QLoRA was used for quantization purposes.

Phi-2 model from Huggingface with BitsandBytes support

Preprocessing [optional]

def format_phi2(row):
    question = row['Context']
    answer = row['Response']

#     text = f"[INST] {question} [/INST] {answer}".replace('\xa0', ' ')
    text = f"### Question: {question}\n ### Answer: {answer}"

    return text

Training Hyperparameters

LoRA config-

# LoRA attention dimension (int)
lora_r = 64

# Alpha parameter for LoRA scaling (int)
lora_alpha = 16

# Dropout probability for LoRA layers (float)
lora_dropout = 0.05

# Bias (string)
bias = "none"

# Task type (string)
task_type = "CAUSAL_LM"

# Random seed (int)
seed = 33

Phi-2 config-

# Batch size per GPU for training (int)
per_device_train_batch_size = 6

# Number of update steps to accumulate the gradients for (int)
gradient_accumulation_steps = 2

# Initial learning rate (AdamW optimizer) (float)
learning_rate = 2e-4

# Optimizer to use (string)
optim = "paged_adamw_8bit"

# Number of training epochs (int)
num_train_epochs = 4

# Linear warmup steps from 0 to learning_rate (int)
warmup_steps = 10

# Enable fp16/bf16 training (set bf16 to True with an A100) (bool)
fp16 = True

# Log every X updates steps (int)
logging_steps = 100

#L2 regularization(prevents overfitting)
weight_decay=0.0

#Checkpoint saves
save_strategy="epoch"

BnB config

# Activate 4-bit precision base model loading (bool)
load_in_4bit = True

# Activate nested quantization for 4-bit base models (double quantization) (bool)
bnb_4bit_use_double_quant = True

# Quantization type (fp4 or nf4) (string)
bnb_4bit_quant_type = "nf4"

# Compute data type for 4-bit base models
bnb_4bit_compute_dtype = torch.bfloat16

Results

Training loss: 2.229 Validation loss: 2.223

More Information [optional]

Phi-2

Model Card Authors [optional]

YuvrajSingh9886

This mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

<!-- ad-footer -->

---

Released freely — support more experiments like it:

More: https://huggingface.co/YuvrajSingh9886