metadata

license: apache-2.0
datasets:
  - JetBrains/KStack
results:
  - task:
      type: text-generation
    dataset:
      name: MultiPL-HumanEval (Kotlin)
      type: openai_humaneval
    metrics:
      - name: pass@1
        type: pass@1
        value: 29.19
tags:
  - code

Model description

This is a repository for the CodeLlama-7b model fine-tuned on the KStack dataset with rule-based filtering, in the Hugging Face Transformers format. KStack is the largest collection of permissively licensed Kotlin code, and so the model is fine-tuned to work better with Kotlin code.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load pre-trained model and tokenizer
model_name = 'JetBrains/CodeLlama-7B-KStack'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to('cuda')

# Create and encode input
input_text = """\
This function takes an integer n and returns factorial of a number:
fun factorial(n: Int): Int {\
"""
input_ids = tokenizer.encode(
    input_text, return_tensors='pt'
).to('cuda')

# Generate
output = model.generate(
    input_ids, max_length=60, num_return_sequences=1, 
    pad_token_id=tokenizer.eos_token_id,
)

# Decode output
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
print(generated_text)

As with the base model, we can use FIM. To do this, the following format must be used:

'<PRE> ' + prefix + ' <SUF> ' + suffix + ' <MID>'

Training setup

The model was trained on one A100 GPU with following hyperparameters:

Hyperparameter	Value
`warmup`	5%
`max_lr`	1e-6
`num_epochs`	1
'attention_dropout'	0.1
`scheduler`	cosine
`total_batch_size`	128 (~65K tokens per step)
`num_epochs`	1

More details about fine-tuning can be found in the technical report (coming soon!).

Fine-tuning data

For tuning the model, we used the KStack dataset, the largest collection of permissively licensed Kotlin code. To increase the quality of the dataset and filter out outliers, such as homework assignments, we filter out the dataset entries according to the following rules:

We filter out files, which belong to low-popular repos (the sum of stars and forks is less than 6)
Next, we filter out files, which belong to repos with less than 5 Kotlin files
Finally, we remove files which have fewer than 20 SLOC

We clean the content of the remaining dataset entries according to the following rules:

We remove all non-ASCII entries
We remove all package lines, such as package kotlinx.coroutines.channels
We remove half of the import lines

We removed half of the imports to avoid potential hallucinations by the model, where it might attempt to import unnecessary libraries. Additionally, packages were removed because this information is only useful at the project level and may introduce additional noise during the learning process.

Evaluation

For evaluation, we used the Kotlin HumanEval dataset, which contains all 161 tasks from HumanEval translated into Kotlin by human experts. You can find more details about the pre-processing necessary to obtain our results, including the code for running, on the datasets's page.

Here are the results of our evaluation:

Model name	Kotlin HumanEval Pass Rate
`CodeLlama-7B`	26.09
`CodeLlama-7B-KStack`	29.19

Ethical Considerations and Limitations

CodeLlama-7B-KStack is a new technology that carries risks with use. The testing conducted to date has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, CodeLlama-7B-KStack's potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate or objectionable responses to user prompts. The model was fine-tuned on a specific data format (Kotlin tasks), and deviation from this format can also lead to inaccurate or undesirable responses to user queries. Therefore, before deploying any applications of CodeLlama-7B-KStack, developers should perform safety testing and tuning tailored to their specific applications of the model.