README.md

---
license: apache-2.0
datasets:
- qingy2024/QwQ-Distill-Data
- AI-MO/NuminaMath-TIR
language:
- en
base_model:
- Qwen/Qwen2-1.5B-Instruct
pipeline_tag: text-generation
library_name: transformers
tags:
- text-generation-inference
- general-purpose
- math
- code
---

![ToI.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/SFtFiJzRxVEMY7nCqquQM.png)

# **OpenRHO-2B-Thinker**

> **OpenRHO-2B-Thinker** is a **general-purpose reasoning model** designed to enhance the cognitive abilities of **edge-deployed large language models (LLMs)** through **reinforcement learning (RL)**. Fine-tuned from **Qwen2-1.5B-Instruct** using the **QwQ distill dataset**, it delivers refined improvements in logical reasoning, structured problem-solving, and lightweight coding — making it highly efficient for **resource-constrained environments**.

## **Key Improvements**

1. **Advanced Reasoning via RL**:
   Built to support symbolic reasoning, logical deduction, and structured problem-solving with high efficiency — specifically optimized for real-time use on edge systems.

2. **Compact Coding Assistant**:
   Enhanced understanding of multiple programming paradigms and syntax across Python, JavaScript, C++, and more. Supports in-situ code generation and debugging for embedded coding scenarios.

3. **Error Detection & Correction**:
   Identifies logic errors, malformed data structures (e.g., JSON, XML), and provides corrections quickly — with lightweight inference and minimal latency.

4. **Instruction Following & Precision**:
   Tuned to follow multi-step instructions with improved contextual memory, offering consistent and precise responses across a variety of prompt types.

5. **Extended Context Compatibility**:
   Maintains support for **128K token inputs** and **8K token outputs**, while remaining lean enough for real-time edge usage with low power consumption.

## **Quickstart with Transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/OpenRHO-2B-Thinker"

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

prompt = "What is a generator function in Python? Explain with an example."
messages = [
    {"role": "system", "content": "You are a helpful and concise AI assistant skilled in programming and reasoning."},
    {"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**

1. **Edge LLM Applications**:
   Built for embedded AI agents, mobile inference, and low-latency chatbots on constrained hardware.

2. **General-Purpose Reasoning**:
   Effective for real-time logical reasoning, structured deduction, and lightweight problem-solving tasks in everyday applications.

3. **Educational & Programming Tools**:
   Helpful for teaching programming and debugging in interactive, constrained environments (e.g., IoT, robotics kits).

4. **Lightweight Conversational Agents**:
   Enables responsive, intelligent interactions in edge-deployed customer service bots, support kiosks, and automation systems.

5. **Multilingual Mini-NLP Tasks**:
   Supports basic multilingual tasks such as translation, summarization, and information retrieval across multiple languages.

6. **Structured Format Generation**:
   Can generate JSON, Markdown, tables, or tabular outputs in lightweight settings for embedded data workflows.

## **Limitations**

1. **Hardware Requirements (Minimal but Non-Zero)**:
   While designed for edge use, optimal performance still benefits from mid-range NPUs, GPUs, or specialized accelerators.

2. **Knowledge Cutoff & Real-Time Awareness**:
   No ability to fetch live data or respond to real-time information beyond its training snapshot.

3. **Limited Creative Output**:
   Less effective for creative writing, abstract thinking, or tasks requiring deep imagination.

4. **Prompt Sensitivity**:
   Outputs can vary based on prompt clarity; structured prompts yield better, more predictable results.

5. **Inherited Biases**:
   May reflect biases from pretraining data. Use caution in sensitive or high-stakes domains.