---
pipeline_tag: text-generation
inference: true
# widget:
# - text: 'Question: Please write a function in Python that performs bubble sort.\n\nAnswer:'
#   example_title: Bubble sort
#   group: Python
license: apache-2.0
datasets:
# Mentionded in paper
- codeparrot/github-code-clean
- bigcode/starcoderdata
# - Stackexchange
# - CommonCrawl
- open-web-math/open-web-math
- math-ai/StackMathQA
# - Arxiv
# - Wikipedia
# - conceptofmind/FLAN_2022 # Original link is broken, we used IBM's filtered version | Phase 2
- nvidia/HelpSteer
metrics:
- code_eval
library_name: transformers
tags:
- code
model-index:
- name: granite-3b-code-base
  results:
  - task:
      type: text-generation
    dataset:
        type: openai_humaneval # https://arxiv.org/pdf/2107.03374
        name: HumanEval
    metrics:
    - name: pass@1
      type: pass@1
      value: 34.1
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: evalplus/humanevalplus # https://arxiv.org/pdf/2305.01210 https://github.com/evalplus/evalplus
        name: HumanEval+
    metrics:
    - name: pass@1
      type: pass@1
      value: 29.9
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: mbpp # https://arxiv.org/abs/2108.07732
        name: MBPP
    metrics:
    - name: pass@1
      type: pass@1
      value: 36.0
      veriefied: false # Check      
  - task:
      type: text-generation
    dataset:
        type: evalplus/mbppplus # 
        name: MBPP+
    metrics:
    - name: pass@1
      type: pass@1
      value: 45.1
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack 
        name: HumanEvalSynthesis(Python)
    metrics:
    - name: pass@1
      type: pass@1
      value: 36.0
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name: HumanEvalSynthesis(JavaScript)
    metrics:
    - name: pass@1
      type: pass@1
      value: 37.2
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name: HumanEvalSynthesis(Java)
    metrics:
    - name: pass@1
      type: pass@1
      value: 40.9
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name: HumanEvalSynthesis(Go)
    metrics:
    - name: pass@1
      type: pass@1
      value: 26.2
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name: HumanEvalSynthesis(C++)
    metrics:
    - name: pass@1
      type: pass@1
      value: 35.4
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name: HumanEvalSynthesis(Rust)
    metrics:
    - name: pass@1
      type: pass@1
      value: 22.0
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(Python)
    metrics:
    - name: pass@1
      type: pass@1
      value: 25.0
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(JavaScript)
    metrics:
    - name: pass@1
      type: pass@1
      value: 18.9
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(Java)
    metrics:
    - name: pass@1
      type: pass@1
      value: 29.9
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(Go)
    metrics:
    - name: pass@1
      type: pass@1
      value: 17.1
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(C++)
    metrics:
    - name: pass@1
      type: pass@1
      value: 26.8
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalExplain(Rust)
    metrics:
    - name: pass@1
      type: pass@1
      value: 14.0
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(Python)
    metrics:
    - name: pass@1
      type: pass@1
      value: 18.3
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(JavaScript)
    metrics:
    - name: pass@1
      type: pass@1
      value: 23.2
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(Java)
    metrics:
    - name: pass@1
      type: pass@1
      value: 29.9
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(Go)
    metrics:
    - name: pass@1
      type: pass@1
      value: 24.4
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(C++)
    metrics:
    - name: pass@1
      type: pass@1
      value: 16.5 
      veriefied: false # Check
  - task:
      type: text-generation
    dataset:
        type: bigcode/humanevalpack  
        name:  HumanEvalFix(Rust)
    metrics:
    - name: pass@1
      type: pass@1
      value: 3.7
      veriefied: false # Check
---
<!-- 
Granite 3B Code Base

Model Summary: few sentences like starcoder
    - Developers:
    - GH repository:
    - Release date:
    - Lincense:

Usage
Intended use
Generation
Fill-in-the-middle

Training Data

Infrastructure

Limitations

Citation
-->

# Granite 3B Code Base
<!-- ![granite](https://github.com/ibm-granite/granite-code-models/blob/main/figures/granite.png) -->

## Model Summary
**Granite 3B Code Base** is a decoder-only code model designed for code generative tasks (e.g., code generation, code explanation, code fixing). It was trained from scratch on 4 trillion tokens sourced from 116 programming languages, ensuring a comprehensive understanding of programming languages and syntax. 

- **Developers:** IBM Research
- **GitHub Repository:** [ibm-granite/granite-code-models](https://github.com/ibm-granite/granite-code-models)
- **Paper:** [Granite Code Models: A Family of Open Foundation Models
for Code Intelligence](https://)
- **Release Date**: May 6th, 2024
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) license.

## Usage
### Intended use
Prominent enterprise use cases of LLMs in software engineering productivity include code generation, code explanation, code fixing, generating unit tests, generating documentation, addressing technical debt issues, vulnerability detection, code translation, and more. All Granite Code Base models, including the **3B parameters model**, are able to handle these tasks as they were trained on a large amount of code data from 116 programming languages. 

### Generation
This is a simple example of how to use Granite Code Base 3B model.

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" # or "cpu"
model_path = "ibm-granite/granite-3b-code-base"

tokenizer = AutoTokenizer.from_pretrained(model_path)

# drop device_map if running on CPU
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
model.eval()

# change input text as desired
input_text = "def generate():"

# tokenize the text
input_tokens = tokenizer(input_text, return_tensors="pt")

# transfer tokenized inputs to the device
for i in input_tokens:
    input_tokens[i] = input_tokens[i].to(device)

# generate output tokens
output = model.generate(**input_tokens)

# decode output tokens into text
output = tokenizer.batch_decode(output)

# loop over the batch to print, in this example the batch size is 1
for i in output:
    print(output)
```

## Training Data
- **Data Collection and Filtering:** Pretraining code data is sourced from a combination of publicly available datasets (e.g., [GitHub Code Clean](https://huggingface.co/datasets/codeparrot/github-code-clean), [Starcoder data](https://huggingface.co/datasets/bigcode/starcoderdata)), and additional public code repositories and issues from GitHub. We filter raw data to retain a list of 116 programming languages. After language filtering, we also filter out low-quality code. 
- **Exact and Fuzzy Deduplication:** We adopt an aggressive deduplication strategy that includes both exact and fuzzy deduplication to remove documents having (near) identical code content.
- **HAP, PII, Malware Filtering:** We apply a HAP content filter that reduces models' likelihood of generating hateful, abusive, or profane language. We also make sure to redact Personally Identifiable Information (PII) by replacing PII content (e.g., names, email addresses, keys, passwords) with corresponding tokens (e.g., ⟨NAME⟩, ⟨EMAIL⟩, ⟨KEY⟩, ⟨PASSWORD⟩). Moreover, we scan all datasets using [ClamAV](https://www.clamav.net/) to identify and remove instances of malware in the source code.
- **Natural Language Datasets:** In addition to collecting code data for model training, we curate several publicly available high-quality natural language datasets to improve models' proficiency in language understanding and mathematical reasoning. Unlike the code data, we do not deduplicate these datasets.

## Infrastructure
We train the Granite Code models using two of IBM's super computing clusters, namely Vela and Blue Vela, both outfitted with NVIDIA A100 and H100 GPUs respectively. These clusters provide a scalable and efficient infrastructure for training our models over thousands of GPUs.

## Limitations
Large Language Models are often prone to generating incorrect information, typically referred to as hallucinations. **Granite 3B Code Base** model is not the exception in this regard. Even though this model is suited for code-related tasks as it is trained on source code from 116 programming languages, the generated code is not guaranteed to work as intended. It can be inefficient and can also contain bugs or exploits. Moreover, Granite Code Base models are **NOT** instruction-following models. Thus, commands like *"Write a function that computes the square root"* may not work well. The model is best treated as a code completion or code infilling model.

## Citation
```
@misc{granite-models,
  author = {author 1, author2, ...},
  title = {Granite Code Large Language Models: IBM Foundation Models for Code},
  journal = {},
  volume = {},
  year = {2024},
  url = {https://arxiv.org/abs/0000.00000},
}
```