Your AI assistant isn’t confusing customers because it’s too small. It’s confusing them because it answers before it thinks.
If you run a small business in Malaysia, you’ve probably tested an AI chatbot. Maybe you fed your FAQ list into one and connected it to WhatsApp for customer enquiries. The replies came back fast — but sometimes they were confidently wrong. A customer asks about a refund policy, and the bot invents a rule that doesn’t exist. Most Malaysian SME owners quietly delete the bot and go back to replying manually. The problem isn’t the model’s size or your data. It’s that most chatbots are trained to predict the next likely word, not to reason their way to an answer. That distinction is worth your attention, because a new technical tutorial from SupraLabs shows exactly how a small language model can be reshaped to think step-by-step before responding — and why that behaviour matters more than raw size.
What Happened
SupraLabs published a practical guide to building a “reasoning-focused” language model using its reasoning corpus on the Hugging Face Hub. The workflow streams a representative sample of 8,000 records instead of downloading the entire collection, then inspects each row’s token length, source repository, task type, and the ratio of reasoning to final answer (source).
The authors then apply four quality filters. They drop samples outside a training-friendly token budget, remove rows with empty or near-empty reasoning, discard traces where a single line repeats more than 30% of the time, and keep only samples whose reasoning makes up between 15% and 97% of the content (source). What remains is converted into a chat format with explicit
The result is a compact model that produces a visible chain of thought before giving its final answer — the same pattern behind DeepSeek-R1 and other reasoning models, but packaged into something small enough to run on an ordinary laptop.
Why This Matters for Malaysian SMEs
Here’s the translation for your business. When you set up a customer service chatbot, the goal isn’t to mimic a human — it’s to give the right answer the first time. A chatbot that reasons will break a refund question into parts: What did the customer buy? When? What does our policy actually say about this situation? Only then does it answer. That’s the difference between a bot that loses a customer and one that keeps them. Imagine a boutique hotel in Penang using a reasoning chatbot to handle booking enquiries in Manglish. The bot thinks through the booking dates, the room availability, and the deposit policy before replying — instead of hallucinating a discount that wrecks your margins.
The data curation part of the tutorial matters just as much. Those filters aren’t about being strict — they’re about keeping only examples that are complete, balanced, and free of repetition (source). For your business, this is a lesson in building a training set from your own customer conversations. Pull your WhatsApp sales chats, your email responses to complaints, and your team’s best handover notes. Cut the conversations that went nowhere. Keep the ones where a problem was properly diagnosed and solved. The quality of that selection matters more than the volume — most SMEs already have enough good material to make a genuinely useful assistant without gathering years of data.
There’s another angle from the tutorial: the model is tiny. You don’t need a giant cluster of servers to make a useful assistant. If SupraLabs can get reasoning behaviour from a 135M-parameter model, a small business can fine-tune a similarly compact model on its own sales scripts, standard operating procedures, and product knowledge. The model becomes your brand voice — trained on how you actually talk to customers, not on generic internet chatter.
“The models that will serve your customers best aren’t the biggest ones. They’re the ones that take a moment to think before they answer — and the ones trained on your actual business conversations, not on the whole internet.”
The Bigger Picture
Reasoning traces are quickly becoming a commodity. The SupraLabs corpus itself pools
There’s also a practical advantage beyond answer quality: a model that reasons is easier to audit. When a customer gets a wrong answer, you can read the
Finally, note the workflow’s emphasis on streaming and sampling rather than downloading everything (source). That’s a mindset shift for SMEs too: you don’t need all your data loaded into a system on day one. Start with a well-chosen sample, run it through quality filters, test the results on real customer queries, and expand from there. Automation doesn’t have to be a big-bang project.
| Tutorial filter | What it does | Your business equivalent |
|---|---|---|
| Token length range | Keeps samples within a training-friendly budget | Keep customer conversations that are complete but not endless threads |
| Degenerate content filter | Drops empty or near-empty reasoning and answers | Cut chats where no real resolution happened |
| Repetition filter | Removes traces where a line repeats over 30% of the time | Remove responses that just loop through the same script |
| Reasoning ratio filter | Keeps samples that reason but also answer | Keep exchanges that diagnose the problem and then solve it |
The takeaway for you is simple: your next chatbot doesn’t need to be bigger or faster. It needs to think before it speaks, and it needs to be trained on the conversations that represent your business at its best. That combination — a small model, a visible reasoning step, and a well-curated dataset from your own customer history — is exactly what the SupraLabs tutorial demonstrates. For a Malaysian SME, it’s a blueprint that finally makes the “AI assistant” promise feel reachable.
Ready to Streamline Your Operations?
Technology moves fast. Your operations should keep up. AutoRunBiz builds AI systems that run your daily workflows — from WhatsApp order capture to accounting. Book a free 15-min ops audit →