138 lines
4.8 KiB
Markdown
138 lines
4.8 KiB
Markdown
---
|
|
license: apache-2.0
|
|
datasets:
|
|
- Magpie-Align/Magpie-Pro-300K-Filtered
|
|
- mlabonne/FineTome-100k
|
|
- unsloth/OpenMathReasoning-mini
|
|
- prithivMLmods/Grade-Math-18K
|
|
language:
|
|
- en
|
|
base_model:
|
|
- Qwen/Qwen3-0.6B
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
tags:
|
|
- text-generation-inference
|
|
- math
|
|
- code
|
|
- moe
|
|
---
|
|
|
|
|
|
|
|
# Magpie-Qwen-CortexDual-0.6B
|
|
|
|
> **Magpie-Qwen-CortexDual-0.6B** is a specialized, general-purpose model designed for **math**, **code**, and **structured reasoning**. Built with **CortexDual thinking mode**, it dynamically adapts to the complexity of a problem, automatically shifting into a stepwise reasoning mode for intricate logic or math tasks. This 0.6B parameter model leverages **80% of the Magpie Pro 330k dataset** and a modular blend of datasets for general-purpose proficiency and domain versatility.
|
|
|
|
> \[!note]
|
|
> GGUF : [https://huggingface.co/prithivMLmods/Magpie-Qwen-CortexDual-0.6B-GGUF](https://huggingface.co/prithivMLmods/Magpie-Qwen-CortexDual-0.6B-GGUF)
|
|
|
|
---
|
|
|
|
## Key Features
|
|
|
|
1. **Adaptive Reasoning via CortexDual**
|
|
Automatically switches into a deeper thinking mode for complex problems, simulating trace-style deduction for higher-order tasks in math and code.
|
|
|
|
2. **Efficient and Compact**
|
|
At 0.6B parameters, it is optimized for deployment in constrained environments while retaining high fidelity in logic, computation, and structural formatting.
|
|
|
|
3. **Magpie-Driven Data Synthesis**
|
|
Trained using 80% of **Magpie Pro 330k**—a high-quality alignment and reasoning dataset—complemented with curated modular datasets for enhanced general-purpose capabilities.
|
|
|
|
4. **Mathematical Precision**
|
|
Fine-tuned for arithmetic, algebra, calculus, and symbolic logic; ideal for STEM learning platforms, math solvers, and step-by-step tutoring.
|
|
|
|
5. **Lightweight Code Assistance**
|
|
Understands and generates code in Python, JavaScript, and other common languages with contextual accuracy and explanation support.
|
|
|
|
6. **Structured Output Generation**
|
|
Specializes in Markdown, JSON, and table outputs, suitable for technical documentation, instruction generation, and structured reasoning.
|
|
|
|
7. **Multilingual Competence**
|
|
Supports over 20 languages with reasoning and translation support, expanding its reach for global educational and development use.
|
|
|
|
---
|
|
|
|
## Quickstart with Transformers
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_name = "prithivMLmods/Magpie-Qwen-CortexDual-0.6B"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype="auto",
|
|
device_map="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
prompt = "Write a Python function to check if a number is prime. Explain each step."
|
|
|
|
messages = [
|
|
{"role": "system", "content": "You are an AI tutor skilled in both math and code."},
|
|
{"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]
|
|
print(response)
|
|
```
|
|
|
|
---
|
|
|
|
## Demo Inference
|
|
|
|
> [!warning]
|
|
non-thinking (direct, reactive, retrieval-based responses)
|
|
|
|
|
|
|
|
> [!warning]
|
|
thinking (reasoning, planning, deeper analysis)
|
|
|
|
|
|
|
|
|
|
---
|
|
|
|
## Intended Use
|
|
|
|
* General-purpose problem solving in math, logic, and code
|
|
* Interactive STEM tutoring and reasoning explanation
|
|
* Compact assistant for technical documentation and structured data tasks
|
|
* Multilingual applications with a focus on accurate technical reasoning
|
|
* Efficient offline deployment on low-resource devices
|
|
|
|
---
|
|
|
|
## Limitations
|
|
|
|
* Lower creativity and open-domain generation due to reasoning-focused tuning
|
|
* Limited context window size due to compact model size
|
|
* May produce simplified logic paths in highly abstract domains
|
|
* Trade-offs in diversity and expressiveness compared to larger instruction-tuned models
|
|
|
|
---
|
|
|
|
## References
|
|
|
|
1. [Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing](https://arxiv.org/pdf/2406.08464)
|
|
2. [Qwen2.5 Technical Report](https://arxiv.org/pdf/2412.15115)
|
|
3. [YaRN: Efficient Context Window Extension of Large Language Models](https://arxiv.org/pdf/2309.00071) |