We've fine-tuned CodeLlama-34B and CodeLlama-34B-Python on an internal Phind dataset that achieve 67.6% and 69.5% pass@1 on HumanEval, respectively. GPT-4 achieves 67%. We've applied OpenAI's decontamination methodology to our dataset to ensure result validity.

More details can be found on our blog post.

Model Details

This model is fine-tuned from CodeLlama-34B and achieves 67.6% pass@1 on HumanEval.

Dataset Details

We fined-tuned on a proprietary dataset of ~80k high quality programming problems and solutions. This dataset consists of instruction-answer pairs instead of code completion examples, making it structurally different from HumanEval. The Phind models were trained for 2 epochs, for a total of ~160k examples shown. LoRA was not used -- both models are a native finetune. We used DeepSpeed ZeRO 3 and Flash Attention 2 to train these models in three hours on 32 A100-80GB GPUs. We used a sequence length of 4096 tokens.

How to Get Started with the Model

Make sure to install Transformers from the main git branch:

pip install git+https://github.com/huggingface/transformers.git

How to Prompt the Model

Please note that this model is somewhat instruction-tuned, but not chat-tuned.

Do not try to use the Llama chat markup with this model. Instead, simply tell it what you want and add "\n: " at the end of your task.

For example:

Write me a linked list implementation: \n

How to reproduce HumanEval Results

To reproduce our results:

python

from transformers import AutoTokenizer, LlamaForCausalLM
from human_eval.data import write_jsonl, read_problems
from tqdm import tqdm

# initialize the model

model_path = "Phind/Phind-CodeLlama-34B-v1"
model = LlamaForCausalLM.from_pretrained(model_path, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_path)

# HumanEval helper

def generate_one_completion(prompt: str):
    tokenizer.pad_token = tokenizer.eos_token
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)

    # Generate
    generate_ids = model.generate(inputs.input_ids.to("cuda"), max_new_tokens=256, do_sample=True, top_p=0.75, top_k=40, temperature=0.1)
    completion = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    completion = completion.replace(prompt, "").split("\n\n\n")[0]

    return completion

# perform HumanEval
problems = read_problems()

num_samples_per_task = 1
samples = [
    dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"]))
    for task_id in tqdm(problems)
    for _ in range(num_samples_per_task)
]
write_jsonl("samples.jsonl", samples)

# run `evaluate_functional_correctness samples.jsonl` in your HumanEval code sandbox

Bias, Risks, and Limitations

This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.

Training details

Hardware Type: 32x A100-80GB
Hours used: 90 GPU-hours
Cloud Provider: AWS
Compute Region: us-east-1