Hands-On with Qwen3-14B: A Reasoning-Centric LLM for Data Scientists

0.1 Introduction: What If LLMs Could Actually Reason?

As data scientists, we constantly move between raw data and real-world decisions. Whether you’re explaining anomalies, generating insights, or deploying models, reasoning is core to our work. But most large language models (LLMs) are still parrots—pattern matchers without deeper understanding.

Enter Qwen3-14B, Alibaba’s newest open-source model. It’s not just another massive transformer—it’s been designed and instruction-tuned with reasoning and conversation in mind. And thanks to the amazing open-source work by Unsloth, we get a full-featured Colab notebook that lets us try it right now, without needing a GPU cluster.

This post is a walkthrough of that notebook:
Colab Notebook: Qwen3 (14B) - Reasoning & Conversational
We’ll unpack each section, explain how things work, and give you hands-on examples to help you integrate Qwen3 into your own workflow.

Tip

This post is part of the VLM Series. Feedback and questions are welcome!

0.2 Model and Tokenizer Loading

Let’s get started by loading the model and tokenizer using HuggingFace Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "unsloth/qwen2-14b-chat-gptq"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", trust_remote_code=True)

0.3 Example: The Reasoning Difference

To see the power of Qwen3, consider this classic chain-of-thought task:

Prompt:

Alice has 3 apples. She gives 1 to Bob, then buys 5 more. How many apples does she have?

prompt = "Alice has 3 apples. She gives 1 to Bob, then buys 5 more. How many apples does she have?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Expected Output:

Let’s break it down. Alice starts with 3 apples. She gives 1 to Bob → 3 - 1 = 2. She buys 5 more → 2 + 5 = 7. Answer: 7 apples.

0.4 System Prompts and Personality

Qwen3 supports “system prompts” that define tone and behavior:

prompt = "<|im_start|>system\nYou are a sarcastic data science tutor.<|im_end|>\n<|im_start|>user\nWhy is my model overfitting?<|im_end|>"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Output:

Oh, I don’t know… maybe because you fed it every single variable and forgot cross-validation? Classic!

0.5 Fine-Tuning and Customization (Optional)

Unsloth supports fine-tuning with LoRA or QLoRA. You could:

• Feed your company’s docs and fine-tune a chatbot
• Inject private datasets and business-specific reasoning
• Modify for multi-modal pipelines

Not covered directly in the notebook—see the next blog post for a tutorial!

0.6 Performance on Data Science Tasks

Qwen3-14B shines at:

• EDA Explanations
• Math QA
• Prompt Chaining
• Code Review

prompt = "Explain what this code does:\ndf.groupby('region')['sales'].mean().sort_values(ascending=False)"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Answer:

This groups the dataframe df by the ‘region’ column, computes the mean sales for each group, and sorts the results in descending order.

0.7 TL;DR

If you’re a data scientist looking for:

• Open, commercial-friendly LLMs
• Strong reasoning abilities
• Easy deployment via Colab

Qwen3-14B is worth your time.

0.8 Bonus: Prompt Engineering Tips

• Be Explicit: Add “Step-by-step reasoning” to prompts.
• Use System Prompts: Tailor tone and format.
• Limit Token Budget: Keep max tokens reasonable for speed + clarity.

---

0.9 Why Qwen3-14B?

Before diving into code, here’s what makes Qwen3 interesting:

• Size and Performance: 14B parameters—big enough to be powerful, small enough to run locally with quantization.
• Open Weight License: Truly open, including for commercial use.
• Reasoning Optimized: Trained with a focus on multi-step logical tasks, coding, math, and chain-of-thought.

0.10 Setup: Running a 14B Model in Google Colab?

The notebook from Unsloth uses the HuggingFace transformers library, AutoGPTQ, and 4-bit quantized weights. That means we can run Qwen3-14B on a free Colab GPU (ideally a T4 or A100) without melting our RAM.

!pip install --upgrade --quiet transformers accelerate auto-gptq

Important

• Run the cell above in your Colab notebook before anything else.
• For best results, use a GPU runtime (T4 or A100 preferred).

0.11 Next Steps

• Back to VLM Series Overview
• Fine-Tuning Qwen3-14B with Unsloth (next post)
• Try the Colab Notebook

0.12 References

1. Qwen3-14B on HuggingFace
2. Qwen3-14B Paper
3. Unsloth Qwen3-14B Colab Notebook
4. VLM Series Overview

Note

This post is part of the VLM Series. Feedback and questions are welcome!