Omega-Qwen2.5-Coder-3B/README.md

---
license: apache-2.0
tags:
- 'Thinking: Disabled'
- Forge
- code
- mot
- stem
- coder
- trl
datasets:
- prithivMLmods/Open-Omega-Forge-1M
language:
- en
- zh
base_model:
- Qwen/Qwen2.5-Coder-3B-Instruct
pipeline_tag: text-generation
library_name: transformers
---
 
![9.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/RXFgDRiO1VBm6jJcH2TPr.png)

# **Omega-Qwen2.5-Coder-3B**

> **Omega-Qwen2.5-Coder-3B** is a compact and high-efficiency code-focused model fine-tuned on **Qwen2.5-Coder-3B-Instruct**, using the symbolic-rich **Open-Omega-Forge-1M** dataset. Designed specifically for **hard-coded tasks** and deterministic computation, this model runs in a **"thinking-disabled"** mode—delivering precise, structured outputs with minimal hallucination, making it ideal for rigorous coding workflows and embedded logic applications.

> Thinking: Disabled

> \[!note]
> GGUF: [https://huggingface.co/prithivMLmods/Omega-Qwen2.5-Coder-3B-GGUF](https://huggingface.co/prithivMLmods/Omega-Qwen2.5-Coder-3B-GGUF)

## **Key Features**

1. **Purpose-Built for Hard Coding**
   Specially tuned to perform precise, low-level code generation with **minimal reasoning overhead**. Ideal for edge-case algorithms, embedded scripting, and deterministic logic patterns.

2. **Optimized Qwen2.5 Foundation**
   Built on **Qwen2.5-Coder-3B-Instruct**, benefiting from its robust token handling, instruction following, and multilingual code representation.

3. **Backed by Open-Omega-Forge-1M Dataset**
   Trained on a curated mix of code, math, and logic problems focused on **symbolic clarity** and **STEM coherence**, drawn from sources like OpenCodeReasoning, MathX-5M, OpenMathReasoning, and more.

4. **Thinking Disabled Mode**
   The model avoids overgeneralizing or injecting speculative reasoning. It executes tasks **as-is**—perfect for structured prompts, tight constraints, and automation pipelines.

5. **Structured Output Control**
   Outputs in **JSON**, **YAML**, **Python**, **Markdown**, and **LaTeX**, tailored for script generation, data serialization, and scientific formatting.

6. **Efficient 3B Deployment**
   Lightweight and scalable for **mid-tier GPUs**, **offline dev environments**, or **local inference systems**, while maintaining solid performance on symbolic tasks.

---

## **Quickstart with Transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Omega-Qwen2.5-Coder-3B"

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

prompt = "Write a Python function to compute factorial iteratively."

messages = [
    {"role": "system", "content": "You are a deterministic code generator. No assumptions. No extra explanations."},
    {"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=256
)
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]
print(response)
```

---

## **Intended Use**

* Embedded logic and deterministic function generation
* Script automation and toolchain integration
* Codegen under fixed constraints or symbolic inputs
* Lightweight STEM applications on edge devices or offline clusters
* Tools where **"no thinking" = better stability**

## **Limitations**

* Not suitable for high-level reasoning or open-ended thought processes
* General chat performance is minimal by design
* Lacks emotional intelligence or creative composition capability
* Assumes user provides **clear, explicit instructions** for best results