Gemma 2 Scripter

Gemma 2 Scripter is a fine-tuned version of the Gemma 2 2B instruct model designed for generating high-quality YouTube scripts based on provided keywords. It is optimized for text generation tasks, delivering coherent and contextually relevant outputs.

Model Details

Model Name: Sidharthan/gemma2_scripter
Architecture: Causal Language Model
Base Model: Gemma 2 2B
Fine-tuning Objective: Script generation using prompt-based keywords.

How to Use

Installation

Ensure you have the following dependencies installed:

pip install torch transformers peft

Code Sample

from transformers import AutoTokenizer
from peft import AutoPeftModelForCausalLM
import torch

# Load the model and tokenizer
model_name = "Sidharthan/gemma2_scripter"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model = AutoPeftModelForCausalLM.from_pretrained(
    model_name,
    device_map=None,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    trust_remote_code=True,
    low_cpu_mem_usage=True
).to(device)

# Generate a script
def generate_script(prompt):
    formatted_prompt = f"<bos><start_of_turn>keywords\n{prompt}<end_of_turn>\n<start_of_turn>script\n"
    inputs = tokenizer(formatted_prompt, return_tensors="pt")
    inputs = {key: value.to(device) for key, value in inputs.items()}
    
    outputs = model.generate(
        **inputs,
        max_length=1024,
        do_sample=True,
        temperature=0.7,
        top_p=0.95,
        top_k=50,
        repetition_penalty=1.2,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id
    )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()

# Example usage
prompt = "crosshatch waffle texture, dark chocolate, four bar crispy wafers, kat, milk chocolate"
response = generate_script(prompt)
print(f"Generated Script:\n{response}")

Input Format

The model expects prompts in the following format:

<bos><start_of_turn>keywords
<your_keywords_here><end_of_turn>
<start_of_turn>script

Example:

<bos><start_of_turn>keywords
crosshatch waffle texture, dark chocolate, four bar crispy wafers, kat, milk chocolate<end_of_turn>
<start_of_turn>script

Output

The output is a YouTube script generated based on the keywords provided.

Performance

CPU: Slower inference due to computational constraints.
GPU: Optimized for faster inference with FP16 support.

Applications

Generating structured scripts for video content
Keyword-based text generation for creative tasks

Training Details

Training Data

The model was fine-tuned on a custom dataset of YouTube scripts paired with their corresponding keywords.

Training Procedure

Fine-tuning Method: LoRA (Low-Rank Adaptation)
Optimization: AdamW optimizer
Learning Rate: 2e-4
Batch Size: 4
Training Steps: 1000

Limitations

The model's output quality depends on the clarity and relevance of input keywords
May occasionally generate repetitive content
Performance may vary based on hardware capabilities

Citation

If you use this model in your research, please cite:

@misc{gemma2_scripter,
  author = {Sidharthan},
  title = {Gemma 2 Scripter: Fine-tuned YouTube Script Generator},
  year = {2024},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub},
  howpublished = {\url{https://huggingface.co/Sidharthan/gemma2_scripter}}
}

License

This model is released under the MIT License.

Sidharthan
/

gemma2_scripter