StableCode: Stability AI Developer Productivity/Developer

“ “a colorful parrot with glasses typing away at a computer, flat icon, vector” – SDXL 0.9 ” — Stable Diffusion XL

This repository contains Stability AI's ongoing development of the StableCode series of code models and will be continuously updated with new checkpoints. The following provides an overview of all currently available models. More coming soon.

News

January 16, 2024

• Released the next version of Stable Code for developers. Catch the release blog post here.

August 9, 2023

• Released an improved version of StableCode-Completion-Alpha-3B -> StableCode-Completion-Alpha-3B-v1.1. This model is trained on more tokens of the top 5 languages.

August 8, 2023

Released the initial suite of StableCode-Alphas. Catch the release blog post here. Which includes,

• StableCode-Completion-Alpha-3B
• StableCode-Completion-Alpha-3B-4k
• StableCode-Instruct-Alpha-3b

January 16, 2023

Released StableCode-3B. Catch the release blog post here. Which includes,

• StableCode-3B a super strong coding model competitive with code-llama-7b in 3B size.

Models

StableCode-Completion-Alpha suite of Models

• StableCode-Completion-Alpha-3B - StableCode-Completion-Alpha-3B is a 3 billion parameter decoder-only code completion model pre-trained on a diverse set of programming languages that were the top used languages based on the 2023 stackoverflow developer survey with a context length of 16k. Trained on a special augmented version of the starcoder-dataset.
• StableCode-Completion-Alpha-3B-4k - StableCode-Completion-Alpha-3B-4K is a 3 billion parameter decoder-only code completion model pre-trained on diverse set of programming languages that topped the stackoverflow developer survey with a context length of 4k.

Training Details

Following similar work, we use a multi-stage approach to context length extension (Nijkamp et al., 2023), scheduling 390 billion tokens at context length 4096 followed by 100 billion tokens at 16k tokens. We found that sequence length warmup (Li et al., 2022) helped stabilize early spikes during the first ~80 billion tokens of pre-training. However, it was not applied to the final runs due to significant throughput penalties as length shapes grew across the curriculum.

Training Data

The training is done in two stages, initial pretraining with the top 12 languages, which we got inspired by Stackoverflow developer survey, The initial subsets - Java, Javascript, Python, TypeScript, PHP, SQL, Rust, C, MarkDown, Go, C++, Shell.

This is then followed by continued pretraining with top 6 languages to be an expert in those languages - Java, Javascript, Python, C, C++, Go

Evaluation

The following zero-shot evaluations are performed with the awesome BigCode Evaluation Harness,

Name	HuggingFace Name	Type		Context Length	Human-Eval
					pass@1
StableCode-Completion-Alpha-3B	stabilityai/stablecode-completion-alpha-3b	Base	3B	16384	20.18
StableCode-Completion-Alpha-3b-4k	stabilityai/stablecode-completion-alpha-3b-4k	Base	3B	4096	17.68
Stablecode-Instruct-Alpha-3b	stabilityai/stablecode-instruct-alpha-3b	Instruction Tuned	3B	4096	26.89
StableCode-Completion-Alpha-3B v1.1	stablecode-completion-alpha-3b@v1.1	Base	3B	16384	22.06
StableCode-3B	stable-code-3b	Base	3B	16384	32.400

Quickstart

All StableCode models are hosted on the Hugging Face hub. Check out this notebook to run inference with limited GPU capabilities.

Get started on generating code with StableCode-Completion-Alpha by using the following code snippet:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList

tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablecode-completion-alpha-3b")
model = AutoModelForCausalLM.from_pretrained("stabilityai/stablecode-completion-alpha-3b")
model.half().cuda()

class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        stop_ids = set([50278, 50279, 50277, 1, 0])
        return input_ids[0][-1] in stop_ids


prompt = f"import torch\nimport torch.nn as nn"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
tokens = model.generate(
  **inputs,
  max_new_tokens=64,
  temperature=0.7,
  do_sample=True,
  stopping_criteria=StoppingCriteriaList([StopOnTokens()])
)
print(tokenizer.decode(tokens[0], skip_special_tokens=True))

Licenses

• Base model checkpoints (StableCode-Completion-Alpha-3B) are licensed under Apache 2.0

• Instruct-tuned checkpoints (StableCode-Instruct-Alpha-3B) are licensed under StableCode Research License Copyright (c) Stability AI Ltd. All Rights Reserved

• All code in this repository is licensed under the Apache License 2.0 license.