metadata

license: apache-2.0

CodeGen2.5-7B-mono

Title: CodeGen2.5: Small, but mighty

Authors: Erik Nijkamp*, Hiroaki Hayashi*, Yingbo Zhou, Caiming Xiong

(* equal contribution)

Model description

CodeGen2.5 is a family of autoregressive language models for program synthesis.

Building upon CodeGen2, the model is trained on StarCoderData for 1.4T tokens, achieving competitive results compared to StarCoderBase-15.5B with less than half the size.

Like CodeGen2, this model is capable of infilling, and supports multiple programming languages.

We then further train on Python, then on instruction data. We release all the models as follows:

CodeGen2.5-7B-multi: Trained on StarCoderData. Licensed under Apache-2.0.
CodeGen2.5-7B-mono (this repo): Further trained on additional Python tokens. Licensed under Apache-2.0.
CodeGen2.5-7B-instruct: Further trained from CodeGen2.5-7B-mono on instruction data. Research purposes only.

How to use

This model can be easily loaded using the AutoModelForCausalLM functionality.

Pre-requisite

Please install OpenAI tiktoken for the tokenizer.

pip install tiktoken==0.4.0

Causal sampling (code autocompletion)

For regular causal sampling, simply generate completions given the context:

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen25-7b-mono", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Salesforce/codegen25-7b-mono")

text = "def hello_world():"
input_ids = tokenizer(text, return_tensors="pt").input_ids
generated_ids = model.generate(input_ids, max_length=128)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))

Infill sampling

For infill sampling, we follow the CodeGen2 format:

<mask_N>: N-th span to be masked. In practice, use <mask_1> to where you want to sample infill.
<sep>: Separator token between the suffix and the infilled sample. See below.
<eom>: "End-Of-Mask" token that model will output at the end of infilling. You may use this token to truncate the output.

For example, if we want to generate infill for the following cursor position of a function:

def hello_world():
    |
    return name

we construct an input to the model by

Inserting <mask_1> token in place of cursor position
Append <sep> token to indicate the boundary
Insert another <mask_1> to indicate which mask we want to infill.

The final snippet looks as follows:

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen25-7b-mono")
model = AutoModelForCausalLM.from_pretrained("Salesforce/codegen2k-7b-mono")


def format(prefix, suffix):
  return prefix + "<mask_1>" + suffix + "<|endoftext|>" + "<sep>" + "<mask_1>"


prefix = "def hello_world():\n    "
suffix = "    return name"
text = format(prefix, suffix)
input_ids = tokenizer(text, return_tensors="pt").input_ids
generated_ids = model.generate(input_ids, max_length=128)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=False)[len(text):])

You might want to truncate the model output with <eom>.

Evaluation results

We evaluate our models on HumanEval and HumanEval-Infill. Please refer to the blog for more details.

Intended use and limitations

As an autoregressive language model, CodeGen2.5 is capable of extracting features from given natural language and programming language texts, and calculating the likelihood of them. However, the model is intended for and best at program synthesis, that is, generating executable code given English prompts, where the prompts should be in the form of a comment string. The model can complete partially-generated code as well.

BibTeX entry and citation info

Please cite CodeGen2 paper:

@article{Nijkamp2023codegen2,
  title={CodeGen2: Lessons for Training LLMs on Programming and Natural Languages},
  author={Nijkamp, Erik and Hayashi, Hiroaki and Xiong, Caiming and Savarese, Silvio and Zhou, Yingbo},
  journal={arXiv preprint},
  year={2023}
}