---
library_name: transformers
tags:
- text-generation-inference
- Deepseek
- code
- math
- RL
- R1
license: apache-2.0
language:
- en
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
pipeline_tag: text-generation
---

![OR1.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/OVqPEwaNZDvlDxsR0osHk.png)

# **Asterope-21-OpenR1**

> **Asterope-21-OpenR1** is a **distributed reinforcement learning (RL)** fine-tuned model based on **Qwen-1.5B**, purpose-built to enhance **coding proficiency**, **debugging accuracy**, and **step-by-step reasoning** in **software development tasks** across multiple programming languages. Compact yet capable, it's ideal for intelligent coding assistants, developer tools, and embedded reasoning engines.

## **Key Features**

1. **Code-Centric Chain-of-Thought Reasoning**  
   Optimized to generate structured, multi-step solutions for programming problems — including algorithm design, debugging, and code explanation — enabling developers to understand the "why" behind each step.

2. **Distributed Reinforcement Learning Fine-Tuning**  
   Trained with reinforcement learning across distributed environments to reinforce optimal coding strategies and accurate logical reasoning pathways.

3. **Multilingual Programming Support**  
   Supports various programming languages (e.g., **Python**, **JavaScript**, **C++**, **Java**, **Go**) and adapts to a wide range of development contexts from scripting to systems programming.

4. **Lightweight, Developer-Ready (1.5B Parameters)**  
   Designed for low-latency environments like IDE extensions, browser dev tools, and CLI bots, making it both fast and resource-efficient.

## **Quickstart with Transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Asterope-21-OpenR1"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Debug the following Python code:\ndef add(a, b):\n  return a + b\nprint(add(5))"
messages = [
    {"role": "system", "content": "You are a skilled coding assistant capable of reasoning step-by-step to solve software development tasks."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```

## **Intended Use**

- **Code Debugging Assistants**: Identifying, explaining, and fixing bugs with precision.
- **Educational Coding Tools**: Helping users learn how and why code works, with rich step-by-step walkthroughs.
- **Multi-language Code Generation**: Write clean, working code across languages and platforms.
- **Lightweight IDE Integration**: Embed into **editors**, **terminals**, or **web-based environments**.

## **Limitations**

1. **Focused Domain**:  
   Optimized for development workflows. May underperform in creative or non-technical tasks.

2. **Model Scale**:  
   Though efficient, complex multi-file or large-context debugging tasks may benefit from larger models.

3. **RL Bias Toward Code Tasks**:  
   Reinforcement learning favors coding reasoning paths — outputs for general-purpose Q&A may be limited.

4. **Prompt Structure Matters**:  
   More effective when inputs include structured error messages, full code context, or clear questions.