Model Card for Model ID

Finetuned Llama3-8B-Instruct model on https://huggingface.co/datasets/isaacchung/hotpotqa-dev-raft-subset.

Model Details

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

Developed by: Isaac Chung
Language(s) (NLP): [English]
License: [Apache 2.0]
Finetuned from model [optional]: meta-llama/Meta-Llama-3-8B-Instruct

How to Get Started with the Model

Use the code below to get started with the model.

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("isaacchung/llama3-8B-hotpotqa-raft")
model = AutoModelForCausalLM.from_pretrained("isaacchung/llama3-8B-hotpotqa-raft")

Training Details

Training Data

https://huggingface.co/datasets/isaacchung/hotpotqa-dev-raft-subset

Training Procedure

Training Hyperparameters

Model loaded:

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
    quantization_config=bnb_config
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.padding_side = 'right' # to prevent warnings

Training params:

# LoRA config based on QLoRA paper & Sebastian Raschka experiment
peft_config = LoraConfig(
        lora_alpha=128,
        lora_dropout=0.05,
        r=256,
        bias="none",
        target_modules="all-linear",
        task_type="CAUSAL_LM",
)

args = TrainingArguments(
    num_train_epochs=3,                     # number of training epochs
    per_device_train_batch_size=3,          # batch size per device during training
    gradient_accumulation_steps=2,          # number of steps before performing a backward/update pass
    gradient_checkpointing=True,            # use gradient checkpointing to save memory
    optim="adamw_torch_fused",              # use fused adamw optimizer
    logging_steps=10,                       # log every 10 steps
    save_strategy="epoch",                  # save checkpoint every epoch
    learning_rate=2e-4,                     # learning rate, based on QLoRA paper
    bf16=True,                              # use bfloat16 precision
    tf32=True,                              # use tf32 precision
    max_grad_norm=0.3,                      # max gradient norm based on QLoRA paper
    warmup_ratio=0.03,                      # warmup ratio based on QLoRA paper
    lr_scheduler_type="constant",           # use constant learning rate scheduler
)

max_seq_length = 3072 # max sequence length for model and packing of the dataset
 
trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=dataset,
    peft_config=peft_config,
    max_seq_length=max_seq_length,
    tokenizer=tokenizer,
    packing=True,
    dataset_kwargs={
        "add_special_tokens": False,  # We template with special tokens
        "append_concat_token": False, # No need to add additional separator token
    }
)

Speeds, Sizes, Times [optional]

train_runtime: 1148.4436
train_samples_per_second: 0.392
train_steps_per_second: 0.065
train_loss: 0.5639963404337565
epoch: 3.0

Training Loss

{'loss': 1.0092, 'grad_norm': 0.27965569496154785, 'learning_rate': 0.0002, 'epoch': 0.4}                                   
{'loss': 0.695, 'grad_norm': 0.17789314687252045, 'learning_rate': 0.0002, 'epoch': 0.8}
{'loss': 0.6747, 'grad_norm': 0.13655725121498108, 'learning_rate': 0.0002, 'epoch': 1.2}                                   
{'loss': 0.508, 'grad_norm': 0.14653471112251282, 'learning_rate': 0.0002, 'epoch': 1.6}                                    
{'loss': 0.4961, 'grad_norm': 0.14873674511909485, 'learning_rate': 0.0002, 'epoch': 2.0}                                   
{'loss': 0.3509, 'grad_norm': 0.1657964587211609, 'learning_rate': 0.0002, 'epoch': 2.4}                                    
{'loss': 0.3321, 'grad_norm': 0.1634644716978073, 'learning_rate': 0.0002, 'epoch': 2.8}

isaacchung
/

llama3-8B-hotpotqa-raft