StarCoder2-Instruct: Fully Transparent and Permissive Self-Alignment for Code Generation

⭐️ About | 🚀 Quick start | 📚 Data generation | 🧑‍💻 Training | 📊 Evaluation | ⚠️ Limitations

About

We introduce StarCoder2-15B-Instruct-v0.1, the very first entirely self-aligned code Large Language Model (LLM) trained with a fully permissive and transparent pipeline. Our open-source pipeline uses StarCoder2-15B to generate thousands of instruction-response pairs, which are then used to fine-tune StarCoder-15B itself without any human annotations or distilled data from huge and proprietary LLMs.

• Model: bigcode/starcoder2-15b-instruct-v0.1
• Code: bigcode-project/starcoder2-self-align
• Dataset: bigcode/self-oss-instruct-sc2-exec-filter-50k
• Authors: Yuxiang Wei, Federico Cassano, Jiawei Liu, Yifeng Ding, Naman Jain, Harm de Vries, Leandro von Werra, Arjun Guha, Lingming Zhang.

Quick start

Here is an example to get started with StarCoder2-15B-Instruct-v0.1 using the transformers library:

```python import transformers import torch

pipeline = transformers.pipeline( model="bigcode/starcoder2-15b-instruct-v0.1", task="text-generation", torchdtype=torch.bfloat16, devicemap="auto", )

def respond(instruction: str, responseprefix: str) -> str: messages = [{"role": "user", "content": instruction}] prompt = pipeline.tokenizer.applychattemplate(messages, tokenize=False) prompt += responseprefix

teminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("###"),
]

result = pipeline(
    prompt,
    max_length=256,
    num_return_sequences=1,
    do_sample=False,
    eos_token_id=teminators,
    pad_token_id=pipeline.tokenizer.eos_token_id,
    truncation=True,
)
response = response_prefix + result[0]["generated_text"][len(prompt) :].split("###")[0].rstrip()
return response

instruction = "Write a quicksort function in Python with type hints and a 'lessthan' parameter for custom sorting criteria." responseprefix = ""

print(respond(instruction, response_prefix)) ```

Data generation pipeline

Run pip install -e . first to install the package locally. Check seed_gathering for details on how we collected the seeds.

By default, we use in-memory vLLM engine for data generation, but we also provide an option to use vLLM's OpenAI compatible server for data generation.

Set CUDA_VISIBLE_DEVICES=... to specify the GPU devices to use for the vLLM engine.

To maximize data generation efficiency, we recommend invoking the script multiple times with different seed_code_start_index and max_new_data values, each with an vLLM engine running on a separate GPU set. For example, for a 100k seed dataset on a 2-GPU machine, you can have 2 processes each generating 50k samples by setting CUDA_VISIBLE_DEVICES=0 --seed_code_start_index 0 --max_new_data 50000 and CUDA_VISIBLE_DEVICES=1 --seed_code_start_index 50000 --max_new_data 50000.

Training Details

Run pip install -e . first to install the package locally. And install Flash Attention to speed up the training.

Hyperparameters

• Optimizer: Adafactor
• Learning rate: 1e-5
• Epoch: 4
• Batch size: 64
• Warmup ratio: 0.05
• Scheduler: Linear
• Sequence length: 1280
• Dropout: Not applied

Hardware

1 x NVIDIA A100 80GB. Yes, you just need one A100 to finetune StarCoder2-15B!

Script

The following script finetunes StarCoder2-15B-Instruct-v0.1 from the base StarCoder2-15B model. /path/to/dataset.jsonl is the JSONL format of the 50k dataset we generated. You can dump the dataset to JSONL to fit the training script.

Evaluation on EvalPlus, LiveCodeBench, and DS-1000

Check evaluation for more details.

Bias, Risks, and Limitations

StarCoder2-15B-Instruct-v0.1 is primarily finetuned for Python code generation tasks that can be verified through execution, which may lead to certain biases and limitations. For example, the model might not adhere strictly to instructions that dictate the output format. In these situations, it's beneficial to provide a response prefix or a one-shot example to steer the model’s output. Additionally, the model may have limitations with other programming languages and out-of-domain coding tasks.

The model also inherits the bias, risks, and limitations from its base StarCoder2-15B model. For more information, please refer to the StarCoder2-15B model card.