Phind/Phind-CodeLlama-34B-Python-v1

1---2license: llama23model-index:4- name: Phind-CodeLlama-34B-v15  results:6  - task:7      type: text-generation8    dataset:9      type: openai_humaneval10      name: HumanEval11    metrics:12    - name: pass@113      type: pass@114      value: 69.5%15      verified: false16tags:17- code llama18---19 20# **Phind-CodeLlama-34B-Python-v1**21We'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.22 23More details can be found on our [blog post](https://www.phind.com/blog/code-llama-beats-gpt4).24 25## Model Details26This model is fine-tuned from CodeLlama-34B-Python and achieves 69.5% pass@1 on HumanEval.27 28## Dataset Details29We 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.30 31## How to Get Started with the Model32 33Make sure to install Transformers from the main git branch:34 35```bash36pip install git+https://github.com/huggingface/transformers.git37```38 39## How to Prompt the Model40**Please note that this model is somewhat instruction-tuned, but not chat-tuned.**41 42Do 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.43 44For example: 45 46```47Write me a linked list implementation: \n48```49 50## How to reproduce HumanEval Results51 52To reproduce our results:53 54```python55 56from transformers import AutoTokenizer, LlamaForCausalLM57from human_eval.data import write_jsonl, read_problems58from tqdm import tqdm59 60# initialize the model61 62model_path = "Phind/Phind-CodeLlama-34B-v1"63model = LlamaForCausalLM.from_pretrained(model_path, device_map="auto")64tokenizer = AutoTokenizer.from_pretrained(model_path)65 66# HumanEval helper67 68def generate_one_completion(prompt: str):69    tokenizer.pad_token = tokenizer.eos_token70    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)71 72    # Generate73    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)74    completion = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]75    completion = completion.replace(prompt, "").split("\n\n\n")[0]76 77    return completion78 79# perform HumanEval80problems = read_problems()81 82num_samples_per_task = 183samples = [84    dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"]))85    for task_id in tqdm(problems)86    for _ in range(num_samples_per_task)87]88write_jsonl("samples.jsonl", samples)89 90# run `evaluate_functional_correctness samples.jsonl` in your HumanEval code sandbox91```92 93## Bias, Risks, and Limitations94 95<!-- This section is meant to convey both technical and sociotechnical limitations. -->96This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.97 98 99## Training details100 101<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->102 103- **Hardware Type:** 32x A100-80GB104- **Hours used:** 90 GPU-hours105- **Cloud Provider:** AWS106- **Compute Region:** us-east-1