NewHope: Harnessing 99% of GPT-4's Programming Capabilities

We introduce NewHope, a chat model based on llama-2-13b, aiming to provide a strong coding capability. NewHope handle different languages including Python, C++, Java, JavaScript, Go, and more. Preliminary evaluation on HumanEval shows that NewHope possesses 99% of GPT-4's programming capabilities.

Contact: SLAM (SUFE Large AI Model) is a research group at Shanghai University of Finance and Economics. cui.wanyun@sufe.edu.cn

TODO: We will release more evaluatation results and training details later.

Evaluation Results

We evaluated NewHope on HumanEval using the official evaluation script by OpenAI. We compared the Pass@1 metric of NewHope with other models. The results of other models are from PapersWithCode.

Model	Pass@1
GPT-4	67.0
NewHope	66.5
PanGu-Coder2 15B	61.6
WizardCoder 15B	57.3
phi-1 1.3B	50.6
GPT-3.5	48.1
phi-1-small	45.0
PaLM-Coder	36.0
CodeGeeX2-6B	35.9

Model Weights

We have open-sourced the model weights.

Usage

To load the NewHope model using Transformers, use the following code:

import torch
from transformers import LlamaTokenizer, LlamaForCausalLM

base_model = "SLAM-group/NewHope"
tokenizer = LlamaTokenizer.from_pretrained(base_model)
model = LlamaForCausalLM.from_pretrained(base_model, torch_dtype=torch.float16, device_map="auto")
# model.config.use_cache is default to `False`. For inference: `model.config.use_cache = True`

Note: At least Huggingface Transformers 4.31.0 is required to load this model.

You can ask NewHope to generate code with instructions. We provide a simple example of how NewHope model generates code with the specific prompt:

# Suppose required tokenizer and model have already been loaded

instruction = "Write a Python function to tell me what the date is today."
prompt = f"<s> ### Instruction:\n{instruction}\n\n### Response:\n"
inputs = tokenizer(prompt, add_special_tokens=False, return_tensors="pt").to("cuda")
output = model.generate(**inputs, do_sample=True, top_p=0.9, max_new_tokens=2048)[0]
decoded_output = tokenizer.decode(output, skip_special_tokens=True).split("### Response:\n")[-1].strip()
print(decoded_output)

You can also interact with NewHope in a dialog manner with the following prompt:

<s> ### Instruction:\nQ1\n\n### Response:\nA1</s><s> ### Instruction:\nQ2\n\n### Response:\nA2</s>

Evaluation

Local setup

Install HumanEval for evaluation. Details
Install dependencies
```
pip install -r requirements.txt
```

For HumanEval, we use the following prompt:

example_input = 'def is_odd(number: int) -> bool:\n    """ Check whether the given number is odd\n    >>> is_odd(3)\n    True\n    >>> is_odd(6)\n    False\n    """\n'
example_output = 'def is_odd(number: int) -> bool:\n    """ Check whether the given number is odd\n    >>> is_odd(3)\n    True\n    >>> is_odd(6)\n    False\n    """\n    return number % 2 == 1'

task_in_humaneval = "REPLACE `task_in_humaneval` WITH THE SPECIFIC TASK IN HUMANEVAL DATA"

prompt = f"<s> ### Instruction:\nComplete the given function below:\n\n{example_input}\n\n### Response:\n{example_output}</s><s> ### Instruction:\nComplete the given function below:\n\n{task_in_human_eval}\n\n### Response:\n"

To reproduce the results on HumanEval, use the following script:

python complete.py --base_model SLAM-group/NewHope --output_dir output --n_gpu 8

The above script will generate samples.jsonl in output_dir, which can be directly evaluated by HumanEval. Evaluation procedure. We conducted the experiment with fp16 on 8xA800, 80GB GPUs, reaching 66.5% on Pass@1 (v.s. GPT4 67.0%).

Citation

@misc{2023newhope,
    title={NewHope: Harnessing 99% of GPT-4's Programming Capabilities},
    author={Wanyun Cui and Qianle Wang},
    howpublished = https://github.com/SLAM-group/newhope,
    year={2023}
}