stable-code

模型	尺寸	Python	C++	Javascript	Java	PHP	Rust
Stable Code	3B	32.4%	30.9%	32.1%	32.1%	24.2%	23.0%
CodeLLama	7B	30.0%	28.2%	32.5%	31.1%	25.7%	26.3%
Deepseek Coder	1.3B	28.6%	29.2%	28.7%	29.0%	23.6%	18.5%
Wizard Coder	3B	31.6%	25.6%	26.2%	25.8%	25.3%	20.4%
StarCoder	3B	21.6%	19.8%	21.5%	20.5%	19.0%	16.9%
Replit Code V1.5	3B	23.0%	25.9%	26.2%	23.6%	23.2%	21.5%
Deci Coder	1B	19.1%	6.8%	18.4%	16.7%	2.1%	1.7%

模型详细信息

开发机构：Stability AI
模型类型：stable-code 模型是基于 Transformer 解码器架构的自动回归语言模型。
语言：英语、代码
联系方式：有关模型的问题和评论，请发送电子邮件至 [email protected]

模型架构

该模型是一个解码器专用的 Transformer，类似于 LLaMA (Touvron 等人，2023 年) 架构，但做了一些修改。

参数	隐藏尺寸	层数	头数	序列长度
2,796,431,360	2560	32	32	16384

位置嵌入：旋转位置嵌入 (Su 等人，2021 年) 应用于前 25% 的头嵌入维度，以改进吞吐量，遵循 Black 等人 (2022 年) 的做法。
分词器：我们使用的是 GPTNeoX 分词器的修改版本。 NeoX。我们添加了特殊标记来训练填充中间 (FIM) 能力，例如 <FIM_PREFIX> 和 <FIM_SUFFIX>，以及其他特殊标记。

训练

训练数据集

该数据集由在 HuggingFace Hub 上可用的经过过滤的开放源代码大型数据集混合而成：Falcon RefinedWeb 提取 (Penedo 等人，2023 年)，以及 CommitPackFT 和 Github Issues (BigCode.，2023 年)，以及 StarCoder (Li 等人，2023 年)。我们还用来自数学领域的数据补充了我们的训练 (Azerbayev, Zhangir 等人，2023 年和 Yu, Longhui 等人，2023 年)。

训练的排名前 18 种编程语言：- C - CPP - Java - JavaScript - CSS - Go - HTML - Ruby - Rust - Markdown - Shell - Php - Sql - R - Typescript - Python - Jupyter-Clean - RestructuredText

使用和限制

预期用途

该模型旨在用作特定于应用程序的微调的基础基础模型。开发者必须评估和微调模型，以确保其在下游应用程序中安全运行。

限制和偏见

作为基础模型，该模型可能会表现出不可靠、不安全或其他不希望的行为，这些行为必须在部署之前通过评估和微调进行纠正。预训练数据集可能包含攻击性或不适当的内容，即使应用了数据清理过滤器，这些内容也可能反映在模型生成的文本中。我们建议用户在生产系统中使用这些模型时谨慎行事。如果模型不适合您的应用程序，或者用于可能对他人造成故意或无意伤害的任何应用程序，请不要使用这些模型。

参考资料

Hugging Face

Stable Code 3B is a 3 billion parameter Large Language Model (LLM), allowing accurate and responsive code completion at a level on par with models such as Code Llama 7b that are 2.5x larger.

**Key Features**

* **NEW** instruct model `ollama run stable-code`
* Fill in Middle Capability (FIM)
* Supports Long Context, trained with Sequences upto 16,384

![spiderchart](https://github.com/jmorganca/ollama/assets/3325447/6c3de7a5-5e10-4884-81fb-3a1b3f566609)

| Model            | Size | Python | C++  | Javascript | Java | PHP  | Rust |
|------------------|------|--------|------|------------|------|------|------|
| **Stable Code**  | 3B   | 32.4%  | 30.9%| 32.1%      | 32.1%| 24.2%| 23.0%|
| CodeLLama        | 7B   | 30.0%  | 28.2%| 32.5%      | 31.1%| 25.7%| 26.3%|
| Deepseek Coder   | 1.3B | 28.6%  | 29.2%| 28.7%      | 29.0%| 23.6%| 18.5%|
| Wizard Coder     | 3B   | 31.6%  | 25.6%| 26.2%      | 25.8%| 25.3%| 20.4%|
| StarCoder        | 3B   | 21.6%  | 19.8%| 21.5%      | 20.5%| 19.0%| 16.9%|
| Replit Code V1.5 | 3B   | 23.0%  | 25.9%| 26.2%      | 23.6%| 23.2%| 21.5%|
| Deci Coder       | 1B   | 19.1%  | 6.8% | 18.4%      | 16.7%| 2.1% | 1.7% |

## Model Details

* **Developed by**: [Stability AI](https://stability.ai/)
* **Model type**: stable-code models are auto-regressive language models based on the transformer decoder architecture.
* **Language(s)**: English, Code
* **Contact**: For questions and comments about the model, please email `lm@stability.ai`

### Model Architecture

The model is a decoder-only transformer similar to the LLaMA ([Touvron et al., 2023](https://arxiv.org/abs/2307.09288)) architecture with the following modifications:

| Parameters     | Hidden Size | Layers | Heads | Sequence Length |
|----------------|-------------|--------|-------|-----------------|
| 2,796,431,360  | 2560        | 32     | 32    | 16384            |

* **Position Embeddings**: Rotary Position Embeddings ([Su et al., 2021](https://arxiv.org/abs/2104.09864)) applied to the first 25% of head embedding dimensions for improved throughput following [Black et al. (2022)](https://arxiv.org/pdf/2204.06745.pdf).
* **Tokenizer**: We use a modified version of the GPTNeoX Tokenizer.[`NeoX`](https://github.com/EleutherAI/gpt-neox). We add special tokens to train for Fill in the Middle (FIM) capabilities like `<FIM_PREFIX>` and `<FIM_SUFFIX>` along with other special tokens.

## Training

### Training Dataset

The dataset is comprised of a filtered mixture of open-source large-scale datasets available on the [HuggingFace Hub](https://hugging-face.cn/datasets): Falcon RefinedWeb extract ([Penedo et al., 2023](https://hugging-face.cn/datasets/tiiuae/falcon-refinedweb)), along with [CommitPackFT](https://hugging-face.cn/datasets/bigcode/commitpackft) and [Github Issues](https://hugging-face.cn/datasets/bigcode/the-stack-github-issues) (BigCode., 2023), and StarCoder ([Li et al., 2023](https://arxiv.org/abs/2305.06161)). We further supplement our training with data from mathematical domains ([Azerbayev, Zhangir, et al., 2023](https://arxiv.org/abs/2310.10631) and, [Yu, Longhui, et al., 2023](https://arxiv.org/abs/2309.12284)).

Top 18 programming languages trained on:
- C
- CPP
- Java
- JavaScript
- CSS
- Go
- HTML
- Ruby
- Rust
- Markdown
- Shell
- Php
- Sql
- R
- Typescript
- Python
- Jupyter-Clean
- RestructuredText

## Use and Limitations

### Intended Use

The model is intended to be used as a foundational base model for application-specific fine-tuning. Developers must evaluate and fine-tune the model for safe performance in downstream applications.

### Limitations and Bias

As a base model, this model may exhibit unreliable, unsafe, or other undesirable behaviors that must be corrected through evaluation and fine-tuning prior to deployment. The pre-training dataset may have contained offensive or inappropriate content, even after applying data cleansing filters, which can be reflected in the model-generated text. We recommend that users exercise caution when using these models in production systems. Do not use the models if they are unsuitable for your application, or for any applications that may cause deliberate or unintentional harm to others.

## References

[Hugging Face](https://hugging-face.cn/stabilityai/stable-code-3b)

粘贴、拖放或单击以上传图片 (.png, .jpeg, .jpg, .svg, .gif)