Edit model card

Stable Code Instruct 3B

Try it out here: https://huggingface.co/spaces/stabilityai/stable-code-instruct-3b


Model Description

stable-code-instruct-3b is a 2.7B billion parameter decoder-only language model tuned from stable-code-3b. This model was trained on a mix of publicly available datasets, synthetic datasets using Direct Preference Optimization (DPO).

This instruct tune demonstrates state-of-the-art performance (compared to models of similar size) on the MultiPL-E metrics across multiple programming languages tested using BigCode's Evaluation Harness, and on the code portions of MT Bench. The model is finetuned to make it useable in tasks like,

  • General purpose Code/Software Engineering like conversations.
  • SQL related generation and conversation.

Please note: For commercial use, please refer to https://stability.ai/membership.


Here's how you can run the model use the model:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("stabilityai/stable-code-instruct-3b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("stabilityai/stable-code-instruct-3b", torch_dtype=torch.bfloat16, trust_remote_code=True)
model = model.cuda()

messages = [
        "role": "system",
        "content": "You are a helpful and polite assistant",
        "role": "user",
        "content": "Write a simple website in HTML. When a user clicks the button, it shows a random joke from a list of 4 jokes."

prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

tokens = model.generate(

output = tokenizer.batch_decode(tokens[:, inputs.input_ids.shape[-1]:], skip_special_tokens=False)[0]

Model Details


Multi-PL Benchmark:

Model Size Avg Python C++ JavaScript Java PHP Rust
Codellama Instruct 7B 0.30 0.33 0.31 0.31 0.29 0.31 0.25
Deepseek Instruct 1.3B 0.44 0.52 0.52 0.41 0.46 0.45 0.28
Stable Code Instruct (SFT) 3B 0.44 0.55 0.45 0.42 0.42 0.44 0.32
Stable Code Instruct (DPO) 3B 0.47 0.59 0.49 0.49 0.44 0.45 0.37

MT-Bench Coding:

Model Size Score
DeepSeek Coder 1.3B 4.6
Stable Code Instruct (DPO) 3B 5.8(ours)
Stable Code Instruct (SFT) 3B 5.5
DeepSeek Coder 6.7B 6.9
CodeLlama Instruct 7B 3.55
StarChat2 15B 5.7

SQL Performance

Model Size Date Group By Order By Ratio Join Where
Stable Code Instruct (DPO) 3B 24.0% 54.2% 68.5% 40.0% 54.2% 42.8%
DeepSeek-Coder Instruct 1.3B 24.0% 37.1% 51.4% 34.3% 45.7% 45.7%
SQLCoder 7B 64.0% 82.9% 74.3% 54.3% 74.3% 74.3%

How to Cite

      title={Stable Code 3B},
      author={Phung, Duy, and Pinnaparaju, Nikhil and Adithyan, Reshinth and Zhuravinskyi, Maksym and Tow, Jonathan and Cooper, Nathan}
Downloads last month
Model size
2.8B params
Tensor type

Spaces using stabilityai/stable-code-instruct-3b 10

Collection including stabilityai/stable-code-instruct-3b

Evaluation results