README.md

---
license: other
license_name: inf
license_link: https://huggingface.co/infly/OpenCoder-1.5B-Base/blob/main/LICENSE
language:
- en
- zh
base_model: infly/OpenCoder-1.5B-Base
pipeline_tag: text-generation
library_name: transformers
tags:
- code
---

<h1 align="center">
  <br>
    OpenCoder-1.5B-Base-32K-via-16K
  <br>
</h1>

<p align="center">
  <a href="https://github.com/sapromak/adaptive-code-completion">Home Page</a> •
  <a href="https://huggingface.co/collections/sapromak/repository-level-pre-trained-opencoder-684206bfc99d48a7e94c0789">Collection</a> •
  <a href="https://openreview.net/forum?id=t9RN9WX4Ic">Paper</a> •
  <a href="https://github.com/sapromak/adaptive-code-completion/blob/main/thesis.pdf">Thesis</a>
</p>

## Description

This model is derived from [OpenCoder-1.5B-Base](https://huggingface.co/infly/OpenCoder-1.5B-Base) by applying an additional context extension fine-tuning with an Adjustment of the Base Frequency parameter of RoPE from 10,000 to 500,000. The number of optimization steps is 512 with a batch size of 128 on sequences of 16,384 length. The repository context is composed based on the _Path Distance_ heuristics, more details on which and other aspects, including all code used, can be found on the [Home Page](https://github.com/sapromak/adaptive-code-completion) of the project. Note that this model is created with the intent to answer specific research questions and __not__ to gain the maximum possible performance on the repository-level code completion setup. Consider it more as a baseline.

The associated research was initialized and conducted by the [JetBrains Research](https://huggingface.co/JetBrains-Research) association.

<div align="center">
  <img src="https://github.com/sapromak/adaptive-code-completion/blob/main/paper/figures/compilation/beyond-training-window/beyond-training-window-inproject.svg?raw=true" width="100%" alt="Performance" />
  <p>Exact Match on the <em>inproject</em> lines of the <em>large-context</em> subset of the <a href="https://huggingface.co/datasets/JetBrains-Research/lca-project-level-code-completion">Project-Level Code Completion task</a> from the <a href="https://arxiv.org/abs/2406.11612">Long Code Arena benchmark</a>. This checkpoint (solid orange curve) demonstrates its best performance at a context length of 32,768. "1K" refers to 1,024 tokens. The star markers denote the context length used during the repository-level pre-training stage.
</div>

## Quickstart

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

model_name = "sapromak/OpenCoder-1.5B-Base-32K-via-16K"
tokenizer_name = "infly/OpenCoder-1.5B-Base"

model = AutoModelForCausalLM.from_pretrained(model_name,
                                             torch_dtype=torch.bfloat16,
                                             device_map="auto",
                                             trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, trust_remote_code=True)

inputs = tokenizer("# write a quick sort algorithm", return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_new_tokens=256)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)
```