Phi-2 is a Transformer with 2.7 billion parameters. It was trained using the same data sources as Phi-1.5, augmented with a new data source that consists of various NLP synthetic texts and filtered websites (for safety and educational value). When assessed against benchmarks testing common sense, language understanding, and logical reasoning, Phi-2 showcased a nearly state-of-the-art performance among models with less than 13 billion parameters.

Training and evaluation data 📚

CodeAlpaca_20K: contains 20K instruction-following data used for fine-tuning the Code Alpaca model.

Training procedure

The following bitsandbytes quantization config was used during training:

quant_method: bitsandbytes
load_in_8bit: True
load_in_4bit: False
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: fp4
bnb_4bit_use_double_quant: False
bnb_4bit_compute_dtype: float32

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2.5e-05
train_batch_size: 4
eval_batch_size: 8
seed: 66
gradient_accumulation_steps: 32
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
0.7631	0.36	50	0.7174
0.6735	0.71	100	0.6949
0.696	1.07	150	0.6893
0.7861	1.42	200	0.6875
0.7346	1.78	250	0.6867

HumanEval results 📊

WIP

Example of usage 👩‍💻

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mrm8488/phi-2-coder"

tokenizer = AutoTokenizer.from_pretrained(model_id, add_bos_token=True, trust_remote_code=True, use_fast=False)

model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.float16, device="auto")

def generate(
        instruction,
        max_new_tokens=128,
        temperature=0.1,
        top_p=0.75,
        top_k=40,
        num_beams=2,
        **kwargs,
):
    prompt = "Instruct: " + instruction + "\nOutput:"
    print(prompt)
    inputs = tokenizer(prompt, return_tensors="pt")
    input_ids = inputs["input_ids"].to("cuda")
    attention_mask = inputs["attention_mask"].to("cuda")
  
    with torch.no_grad():
        generation_output = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=max_new_tokens,
            eos_token_id = tokenizer.eos_token_id,
            use_cache=True,
            early_stopping=True
        )
    output = tokenizer.decode(generation_output[0])
    return output.split("\nOutput:")[1].lstrip("\n")

instruction = "Design a class for representing a person in Python."
print(generate(instruction))

How to use with MLX.

bash

# Install mlx, mlx-examples, huggingface-cli
pip install mlx
pip install huggingface_hub hf_transfer
git clone https://github.com/ml-explore/mlx-examples.git

# Download model
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli download --local-dir phi-2-coder mrm8488/phi-2-coder

# Run example
python mlx-examples/llms/phi2.py --model-path phi-2-coder --prompt "Design a class for representing a person in Python" 

Citation

@misc {manuel_romero_2023,
	author       = { {Manuel Romero} },
	title        = { phi-2-coder (Revision 4ae69ae) },
	year         = 2023,
	url          = { https://huggingface.co/mrm8488/phi-2-coder },
	doi          = { 10.57967/hf/1518 },
	publisher    = { Hugging Face }
}