PhiMerge-2.7B-Dare-daser

PhiMerge-2.7B-Dare-daser is a mixture of two techniques that are LaserQlora and Dora. This model is a DPO fine-tuned of johnsnowlabs/PhiMerge-2.7B-Dare using the argilla/distilabel-capybara-dpo-7k-binarized preference dataset. The model has been trained on top 16 projections (q_proj, k_proj, v_proj) based on snr values. This model has been trained for 1080 steps.

🏆 Evaluation results

Coming Soon

Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "johnsnowlabs/PhiMerge-2.7B-Dare-daser"
messages = [{"role": "user", "content": "Explain what is Machine learning."}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-04
• train_batch_size: 1
• eval_batch_size: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: paged_adamw_32bit
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 1080

LoRA Config

• lora_r: 16
• lora_alpha: 32
• lora_dropout: 0.05
• peft_use_dora: true

Framework versions

• Transformers 4.38.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.17.0
• Tokenizers 0.15.0