Llama3-8B-SuperNova-Spectrum-Hermes-DPO

This model is a DPO fine-tuned version of my DARE_TIES merged Model yuvraj17/Llama3-8B-SuperNova-Spectrum-dare_ties on the yuvraj17/chatml-OpenHermes2.5-dpo-binarized-alpha-2k dataset.

DPO (Direct Preference Optimization):

Direct Preference Optimization (DPO) is a fine-tuning technique that focuses on aligning a model's responses with human preferences or ranking data without requiring reinforcement learning steps, like in RLHF.

DPO vs RLHF Reference

Training:

• Trained on 1x A40s (48GB VRAM) using the HuggingFace TRL.
• QLoRA(4-bit precision) for 1 epoch

  # LoRA configuration
  peft_config = LoraConfig(
      r=32,
      lora_alpha=16,
      lora_dropout=0.05,
      bias="none",
      task_type="CAUSAL_LM",
      target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
  )
  

Training Params

The following hyperparameters were used during training:

• learning_rate: 5e-05
• beta=0.1
• num_devices: 1
• gradient_accumulation_steps: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 1

Training Time = 1:57:00 hours

Weight & Biases Report

Report-Link

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "yuvraj17/Llama3-8B-SuperNova-Spectrum-Hermes-DPO"
messages = [{"role": "user", "content": "What is a large language model?"}]

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"])

🏆 Evaluation Scores

Coming Soon