README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
language:
  - en
tags:
  - reasoning
  - math
  - chain-of-thought
  - qwen2.5
  - gguf
  - text-generation
  - conversational
pipeline_tag: text-generation
library_name: transformers
---

<div align="center">

![1789911910f175](https://cdn-uploads.huggingface.co/production/uploads/6a5f8c87a38cac087c2c6c05/zZ-C-s-RZWV6sxVhj9RBe.png)

🐍 Boomslang (3B Reasoning & Math Engine)

**A compact, high-efficiency 3-billion-parameter model fine-tuned for deep chain-of-thought mathematical reasoning, logic, and algebra—without sacrificing natural conversational ability.**

[![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Creator-Monster--Code-blue)](https://huggingface.co/Monster-Code)
[![License](https://img.shields.io/badge/License-Apache_2.0-green.svg)](https://opensource.org/licenses/Apache-2.0)
[![GGUF Available](https://img.shields.io/badge/GGUF-Included-purple.svg)](https://huggingface.co/Monster-Code/Boomslang/blob/main/boomslang-3b-qwen.gguf)

---

### ❤️ If you find Boomslang useful, please hit the **Like** button at the top of this page and [Follow @Monster-Code](https://huggingface.co/Monster-Code) for more open-weights AI releases!

---

</div>

## 💡 What is Boomslang?

Most small models (1B–3B parameters) struggle with two extremes: they are either polite chatbots that completely hallucinate basic arithmetic, or narrow math models that forget how to hold a conversation and start writing unprompted proofs when you simply say "Hi."

**Boomslang** was trained to bridge that gap.

Starting from the strong foundation of **`Qwen/Qwen2.5-3B-Instruct`**, Boomslang was post-trained on an **NVIDIA RTX PRO 6000 Blackwell** across a curated ~17,500-sample reasoning mixture:
1. **DeepSeek-R1 Distilled Proofs (`open-r1/OpenR1-Math-220k`):** Teaches the network an internal self-reflection loop (`<think> ... </think>`) to break down complex algebraic expressions, geometry, and multi-step deduction before committing to an answer.
2. **Step-by-Step Arithmetic Rigor (`openai/gsm8k`):** Calibrates attention heads on strict order-of-operations arithmetic and unambiguous answer derivation.

The result is an edge-friendly 3B model that works through tricky algebra and word puzzles methodically, but still greets you warmly and follows instructions when you just want to talk.

---

## 📦 What's Inside This Repository?

* **Single Standalone `model.safetensors`:** No multi-part file splits. The entire 3-billion parameter model is packed into a single, clean ~6.1 GB file.
* **Pre-Converted GGUF (`boomslang-3b-qwen.gguf`):** Directly ready for **Ollama**, **LM Studio**, and **llama.cpp** on your local machine (MacBook, laptop, or home GPU).
* Full configuration and tokenizer files for immediate `transformers` plug-and-play.

---

## ⚡ Quickstart: Python & Transformers

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

MODEL_ID = "Monster-Code/Boomslang"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    device_map="auto"
)

# Standard ChatML format
messages = [
    {
        "role": "system",
        "content": "You are Boomslang, a helpful and precise reasoning assistant. Reason step-by-step before providing your final answer."
    },
    {
        "role": "user",
        "content": "A box has 36 chocolates. Tom eats 1/4 of them, and Sarah eats 9. How many chocolates are left in the box?"
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

with torch.no_grad():
    model.generate(
        **inputs,
        streamer=streamer,
        max_new_tokens=1024,
        temperature=0.6,
        top_p=0.9
    )
```

---

## 💻 Running Locally with Ollama / LM Studio

### 1. Using Ollama:
Create a `Modelfile` in the same directory as the downloaded `.gguf`:
```dockerfile
FROM ./boomslang-3b-qwen.gguf

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.6
```

Build and run:
```bash
ollama create boomslang -f Modelfile
ollama run boomslang
```

### 2. Using LM Studio:
1. Download **`boomslang-3b-qwen.gguf`** from the [Files and versions tab](https://huggingface.co/Monster-Code/Boomslang/tree/main).
2. Drop it into your local LM Studio models folder.
3. Start chatting with GPU offloading enabled!

---

## 🧠 Sample Interaction

**User:**
> *"Solve for x: 5x - 10 = 25. Show your work."*

**Boomslang:**
```text
<think>
The equation is 5x - 10 = 25.
Our goal is to isolate the variable x on one side.

Step 1: Add 10 to both sides to cancel out the -10 constant:
5x - 10 + 10 = 25 + 10
5x = 35

Step 2: Divide both sides by the coefficient 5:
5x / 5 = 35 / 5
x = 7

Verification:
Substitute x = 7 into the original expression:
5(7) - 10 = 35 - 10 = 25.
Both sides match.
</think>

The answer is x = 7.
```

---

## 🛠️ Training Details

* **Base Architecture:** Qwen2.5 3B (Decoder-only Transformer)
* **Hardware:** NVIDIA RTX PRO 6000 Blackwell Server Edition
* **Precision:** BF16 Native Mixed Precision with Fused AdamW
* **Effective Batch Size:** 32 (8 per device × 4 gradient accumulation steps)
* **Learning Rate:** 1.5e-4 with dynamic linear warmup
* **Label Masking:** Dynamic batch padding via `DataCollatorForSeq2Seq` with `-100` masking to guarantee loss is never calculated on padding noise

---

## 🤝 Community & Support

* 👤 **Creator:** [Monster-Code](https://huggingface.co/Monster-Code)
* 💬 Have suggestions, evaluation runs, or dataset ideas? Leave a note in the **Discussions** tab!
* ⭐ **If Boomslang helps your workflow, please consider starring/liking the repository!**