mobiuslabsgmbh/aanaphi2-v0.1

aanaphi2-v0.1 is a finetuned (SFT + DPO) chat model based on Microsoft's Phi-2 base model (2.8B parameters).

Performance

Models	phi-2	aanaphi2-v0.1
ARC (25-shot)	61.09	63.74
HellaSwag (10-shot)	75.11	78.30
MMLU (5-shot)	58.11	57.70
TruthfulQA-MC2	44.47	51.56
Winogrande (5-shot)	74.35	73.40
GSM8K (5-shot)	54.81	58.61
Average	61.33	63.89

Installation

Make sure you have the latest version of the transformers library:

pip install pip --upgrade && pip install transformers --upgrade

Basic Usage

#Load model
import transformers, torch

#GPU runtime
device        = 'cuda'
compute_dtype = torch.float16

##CPU runtime
#device        = 'cpu'
#compute_dtype = torch.float32

cache_path    = ''
model_id      = "mobiuslabsgmbh/aanaphi2-v0.1"
model         = transformers.AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=compute_dtype, 
                                                                  cache_dir=cache_path,
                                                                  device_map=device)
tokenizer     = transformers.AutoTokenizer.from_pretrained(model_id, cache_dir=cache_path)

#Set Prompt format
instruction_template = "### Human: "
response_template    = "### Assistant: "
def prompt_format(prompt):
    out = instruction_template + prompt + '\n' + response_template
    return out
model.eval();

@torch.no_grad()
def generate(prompt, max_length=1024):
    prompt_chat = prompt_format(prompt)
    inputs      = tokenizer(prompt_chat, return_tensors="pt", return_attention_mask=True).to(device)
    outputs     = model.generate(**inputs, max_length=max_length, eos_token_id= tokenizer.eos_token_id) 
    text        = tokenizer.batch_decode(outputs[:,:-1])[0]
    return text

#Generate
print(generate('If A+B=C and B=C, what would be the value of A?'))