The Micro-Based model is the smallest model available in Minibase. It is designed for lightweight tasks and optimized for mobile and edge deployment. While it does not provide deep reasoning or long, dynamic conversations, it performs well on simple, repeatable tasks where responses are short and predictable.
Typical use cases include:
Classification: labeling text or short inputs with predefined categories.
Extraction: pulling out keywords or structured values from text.
Simple Q&A: answering direct factual questions with short responses.
Instruction Following (basic): handling narrow, repeatable tasks reliably.
This model is best when efficiency and small memory footprint are more important than conversational depth or creativity.
Model Specs
# of Parameters: ~135M
Format Type:
causal_lm
Download Size: ~135 MB (estimate)
Max Sequence Length: 1,024 tokens
Context Window: 1,024 tokens
The compact size makes it fast and efficient, ideal for edge devices and resource-limited environments.
How to Use
Step 1: Choose the model
From the Models tab, select the Micro-Based model when creating a new fine-tuned model.
Step 2: Prepare your dataset
Use Instruction, Input, and Response fields, keeping examples short and consistent.
Example:
Instruction: “Classify sentiment as Positive or Negative”
Input: “I love this product!”
Response: “Positive”
Step 3: Train your model
Upload in CSV, Excel, JSON, or JSON-L format. All files are standardized to JSON-L.
Start fine-tuning; training uses LoRA for efficient adaptation.
Step 4: Test in the browser
On the Model Details page, you can chat with your model directly.
Adjust Temperature (randomness) and Max Tokens (length) to evaluate its behavior.
Step 5: Deploy or download
Download in GGUF format for local or mobile use.
Or deploy instantly with Minibase Cloud to access it via APIs.
Tips & Best Practices
Keep tasks narrow: This model works best when it only has to do one specific job repeatedly.
Use small, clean datasets: Consistency matters more than size for micro tasks.
Short outputs are ideal: Design tasks around labels, keywords, or short responses rather than long-form text.
Quantization options:
High (Q8_0): best quality, near-lossless.
Medium (Q4_K_M): balanced, recommended for most users.
Low (Q4_K_S): smallest, fastest, optimized for mobile/edge.
Troubleshooting
Responses are inconsistent
Responses are inconsistent
Ensure training examples are uniform and tightly scoped. Variability can confuse smaller models.
The model doesn’t give long answers
The model doesn’t give long answers
The Micro-based model isn’t designed for that. Fix: Use the Task-based model or a larger model better suited for instruction following and reasoning.
The Micro-based model isn’t designed for that. Fix: Use the Task-based model or a larger model better suited for instruction following and reasoning.
Training still feels slow
Training still feels slow
While smaller models train faster, dataset size matters. Try a smaller dataset for testing, then scale up once satisfied with the model's performance.