OpenBezoar-HH-RLHF-DPO

The OpenBezoar-HH-RLHF-DPO is an LLM that has been fine tuned for human preferences alignment using Direct Preference Optimization (DPO), on top of OpenBezoar-HH-RLHF-SFT model on a subset of Anthropic's HH-RLHF Dataset.

Model Details

• Base Model: OpenBezoar-HH-RLHF-SFT
• Dataset used for SFT: First 100K examples of the HH-RLHF dataset
• Alignment Method: Direct Preference Optimization (DPO)
• Epochs: 1

Model Description

OpenBezoar-HH-RLHF-SFT is an LLM that is built upon the OpenLLaMA 3B v2 architecture. This model has been fine-tuned for human preferences alignment using DPO. Alignment has been performed on top of the OpenBezoar-HH-RLHF-SFT model. For more information please refer to our paper.

Model Sources

• Repository: Bitbucket Project
• Paper : Pre-Print

Instruction Format

We follow a modified version of the Alpaca prompt template as shown below. It is important to utilize this template in order to obtain best responses for instruction related tasks.

### System:
Below is an instruction that describes a task, optionally paired with an input that provides further context following that instruction. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response:

Notice that no end-of-sentence (eos) token is being appended.

Note: The system prompt shown in the following figure is the one that the model has been trained on most of the time. However, you may attempt to use any other system prompt that is available in the Orca scheme.

Usage

from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, AutoModelForSeq2SeqLM

checkpoint =  "SurgeGlobal/OpenBezoar-HH-RLHF-DPO"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)

model = AutoModelForCausalLM.from_pretrained(
	checkpoint,
	load_in_4bit=True, # optionally for low resource environments
	device_map="auto"
)

prompt =  """### System:
Below is an instruction that describes a task, optionally paired with an input that provides further context following that instruction. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response:""".format(
	instruction="What is the world state in the year 1597."
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024, do_sample=True)

print(tokenizer.decode(outputs[0]))

Evaluations

Refer to our self-reported evaluations in our paper (Section 4).

Limitations

• The model might not consistently show improved abilities to follow instructions, and it could respond inappropriately or get stuck in loops.
• Although this model is aligned to human preferences and has been evaluated for performance, it is not guaranteed that it will refrain from generating harmful content exclusively.
• Caution is urged against relying on this model for production or adjacent use-cases.

Citation

If you find our work useful, please cite our paper as follows:

@misc{surge2024openbezoar,
      title={OpenBezoar: Small, Cost-Effective and Open Models Trained on Mixes of Instruction Data}, 
      author={Chandeepa Dissanayake and Lahiru Lowe and Sachith Gunasekara and Yasiru Ratnayake},
      year={2024},
      eprint={2404.12195},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Model Authors

Chandeepa Dissanayake, Lahiru Lowe, Sachith Gunasekara, and Yasiru Ratnayake