README.md

---
datasets:
- bigcode/starcoderdata
language:
- code
tags:
- causal-lm
license: cc-by-sa-4.0
---
# `StableCode-Completion-Alpha-3B`

## Model Description

`StableCode-Completion-Alpha-3B` is a 3 billion parameter decoder-only code completion model pre-trained on diverse set of programming languages that topped the stackoverflow developer survey. 

## Usage
The model is intended to do single/multiline code completion from a long context window upto 16k tokens.
Get started generating code with `StableCode-Completion-Alpha-3B` by using the following code snippet:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablecode-completion-alpha-3b")
model = AutoModelForCausalLM.from_pretrained(
  "stabilityai/stablelm-base-alpha-3b-v2",
  trust_remote_code=True,
  torch_dtype="auto",
)
model.cuda()
inputs = tokenizer("import torch\nimport torch.nn as nn", return_tensors="pt").to("cuda")
tokens = model.generate(
  **inputs,
  max_new_tokens=48,
  temperature=0.2,
  do_sample=True,
)
print(tokenizer.decode(tokens[0], skip_special_tokens=True))
```

## Model Details

* **Developed by**: Code.AI Team @ [Stability AI](https://stability.ai/)
* **Model type**: `StableCode-Completion-Alpha-3B` models are auto-regressive language models based on the transformer decoder architecture.
* **Language(s)**: Code
* **Library**: [GPT-NeoX](https://github.com/EleutherAI/gpt-neox)
* **License**: Model checkpoints are licensed under the Creative Commons license ([CC BY-SA-4.0](https://creativecommons.org/licenses/by-sa/4.0/)). Under this license, you must give [credit](https://creativecommons.org/licenses/by/4.0/#) to Stability AI, provide a link to the license, and [indicate if changes were made](https://creativecommons.org/licenses/by/4.0/#). You may do so in any reasonable manner, but not in any way that suggests the Stability AI endorses you or your use.
* **Contact**: For questions and comments about the model, please email `lm@stability.ai`

### Model Architecture

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


* **Decoder Layer**: Parallel Attention and MLP residuals with a single input LayerNorm ([Wang & Komatsuzaki, 2021](https://github.com/kingoflolz/mesh-transformer-jax/tree/master))
* **Position Embeddings**: Rotary Position Embeddings ([Su et al., 2021](https://arxiv.org/abs/2104.09864))
* **Bias**: LayerNorm bias terms only

## Training

`StableCode-Completion-Alpha-3B` is pre-trained using a multi-stage context length extension schedule following similar work ([Nijkamp et al. 2023](https://blog.salesforceairesearch.com/xgen/)); first pre-training at a context length of 4096 for 300 billion tokens, then fine-tuning at a context length of 16384 for another 200B tokens.

### Training Dataset

The first pre-training stage relies on 300B tokens sourced from various top programming languages occuring in the stackoverflow developer survey in the `starcoder-data` dataset. We then finetune it on a longer context augmentation of `starcoder-data` dataset.  

### Training Procedure

The model is pre-trained on the dataset mixes mentioned above in mixed-precision BF16), optimized with AdamW, and trained using the NeoX tokenizer with a vocabulary size of 49k.

* **Software**: We use a fork of gpt-neox ([EleutherAI, 2021](https://github.com/EleutherAI/gpt-neox)) and train under 2D parallelism (Data and Tensor Parallel) with ZeRO-1 ([Rajbhandari et al., 2019](https://arxiv.org/abs/1910.02054v3)) and rely on flash-attention as well as rotary embedding kernels from FlashAttention-2 ([Dao et al., 2023](https://tridao.me/publications/flash2/flash2.pdf))

## Use and Limitations

### Intended Use


### Limitations and bias