Modelpublic

ardaorcun/finetuned_cosmos2603

sourceHugging Faceupdated 2y agoView on Hugging Face

1likes18downloads

Model Card

Model Card for Model ID

This model is a fine-tuned version of YTU's Cosmos GPT2 Language Model. You can check the code from here:<a href="https://github.com/Stealeristaken/Entry-Mid-Level-AI-Projects/blob/main/Fine%20Tuning%20Cosmos%20by%20LoRA%20and%20QLoRA.ipynb">Fine Tuning Cosmos by LoRA and QLoRA</a>

Training Details

The model was fine-tuned using LoRA and QLoRA techniques. Training parameters are defined below.

LoRA configs:

—r=16
—lora_alpha=32
—target_modules=cproj,cfc, gateproj, cproj, c_attn
—lora_dropout=0.05
—bias="lora_only"
—fan_in_fan_out=True
—max_seq_length=512
—use_rslora=True

Train Parameters:

—train_epochs=5
—optim="pagedlion8bit"
—learning_rate=2e-4
—warmup_ratio=0.03
—max_grad_norm=0.3
—lr_scheduler_type="linear"

Training Data

For training, I used Merve's Turkish Instructions Dataset, which you can check here: <a href="https://huggingface.co/datasets/merve/turkish_instructions">Merve's Turkish Instructions Dataset</a>

Instruction template:

python

def format_instruction(sample):
    return f"""Sen cevap vermeyi seven yardımcı bir dil modelisin.
        ### Input:
        {sample["talimat"]}
        
        ### Context:
        {sample[" giriş"]}

        ### Response:
        {sample[" çıktı"]}
    """

Generate Output:

python

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "ardaorcun/finetuned_cosmos2603"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, 
                                             device_map='auto', 
                                             load_in_8bit=True)

sampling_params = dict(do_sample=True, temperature=0.3, top_k=50, top_p=0.9)

pipe = pipeline("text-generation", 
                model=model, 
                tokenizer=tokenizer,
                device_map="auto",
                max_new_tokens=512, 
                return_full_text=True,
                repetition_penalty=1.1
               )

DEFAULT_SYSTEM_PROMPT = "Sen cevap vermeyi seven yardımcı bir dil modelisin.\n"

def format_instruction(sample):
    return f"""{DEFAULT_SYSTEM_PROMPT}
### Input:
{sample["talimat"]}

### Context:
{sample["giriş"]}

### Response:
{sample["çıktı"]}"""

Create Answer:

python

prompt = "your_prompt"
girdi = "your_entry"
instruction = f"""Sen cevap vermeyi seven yardımcı bir dil modelisin.\n### Input:\n{prompt}\n\n### Context:\n{girdi}\n\n### Response:"""
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_length = 2048)
result = pipe(instruction)
print(result[0]['generated_text'][len(instruction):])