LoRA Adapters for Gemma-2-9B-IT on YouTube Titles

These are LoRA adapters for google/gemma-2-9b-it trained on AdamLucek/youtube-titles.

Intended task is to tune Gemma 2 9B to generate YouTube title's more similar to popular YouTubers, data was prepped in the instruction tuned token format.

Intended uses & limitations

See original model page google/gemma-2-9b-it intended usage for details about Gemma 2 9B usage, limitations, and ethical considerations.

Usage

The below code will show you how to load and interface with the LoRA model.

Loading the Model & LoRA Adapters

from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# Load the Pre Trained Model
model = AutoModelForCausalLM.from_pretrained("google/gemma-2-9b-it", 
                                             quantization_config=BitsAndBytesConfig(load_in_8bit=True), 
                                             device_map="auto"
                                             ).eval()
# Load the Tokenizer
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b-it")

# Attach LoRA Adapters to Pre Trained Model
model = PeftModel.from_pretrained(model, "AdamLucek/gemma-2-9b-it-lora-yt-titles", adapter_name="youtube_titles")

Inference

topic = "huggingface AI models"
messages = [
    {"role": "user", "content": f"Create a YouTube title about {topic}"}
]

# Apply chat template and prepare inputs
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}

# Generate outputs
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=True,
    top_p=0.95,
    temperature=0.1,
    repetition_penalty=1.2,
    eos_token_id=tokenizer.eos_token_id
)
  
# Decode outputs
decoded = tokenizer.decode(outputs[0])

print(decoded)

Training procedure

Training hyperparameters

Trained on a single a6000 using the following script

python \
    examples/scripts/sft.py \
    --model_name_or_path="google/gemma-2-9b-it" \
    --dataset_name="AdamLucek/youtube-titles" \
    --dataset_text_field="gemma2_9b_it_format" \
    --per_device_train_batch_size=4 \
    --per_device_eval_batch_size=4 \
    --gradient_accumulation_steps=4 \
    --max_grad_norm=1.0 \
    --learning_rate=5e-5 \
    --weight_decay=0.01 \
    --lr_scheduler_type="cosine" \
    --warmup_ratio=0.1 \
    --report_to="wandb" \
    --bf16 \
    --max_seq_length=2048 \
    --lora_r=16 \
    --lora_alpha=32 \
    --lora_target_modules q_proj k_proj v_proj o_proj \
    --load_in_8bit \
    --use_peft \
    --attn_implementation="eager" \
    --logging_steps=1 \
    --eval_strategy="steps" \
    --eval_steps=200 \
    --save_strategy="steps" \
    --save_steps=250 \
    --output_dir="models/gemma2" \
    --hub_model_id="gemma-2-9b-it-lora-yt-titles" \
    --push_to_hub \
    --num_train_epochs=3

Training results

It achieves the following results on the evaluation set:

• Loss: 2.2142

Framework versions

• PEFT 0.11.1
• Transformers 4.42.3
• Pytorch 2.0.1
• Datasets 2.20.0
• Tokenizers 0.19.1