Overview
The Task-Based model is designed for problems where the input and output are tightly linked and need to follow a specific structure. It is excellent for tasks such as:
Summarization: condensing long passages into concise summaries.
Translation: converting text between supported languages (English ↔ German, French, Spanish, Romanian).
Question Answering: providing direct responses to factual or structured questions.
Text Classification & Sentiment Analysis: assigning labels to text or determining tone.
Paraphrasing & Rewriting: producing alternative versions of text while preserving meaning.
Instruction Following: responding precisely when you want the model to follow a specific prompt format.
This model is especially useful when you want reliable, context-aware completions that follow clear instructions.
Model Specs
# of Parameters: ~248M
Format Type:
task-basedDownload Size: ~248 MB (estimate)
Max Sequence Length: 512 tokens
Context Window: 512 tokens
These specs make the model lightweight enough to run efficiently while still providing high-quality outputs for structured language tasks.
How to Use
Step 1: Choose the model
From the Models tab, select the Task-Based model when creating a new fine-tuned model.
Step 2: Prepare your dataset
Use fields like Instruction, Input, and Response.
Example:
Instruction: “Translate English to German”
Input: “The cat is on the table.”
Response: “Die Katze ist auf dem Tisch.”
Step 3: Train your model
Upload your dataset in CSV, Excel, JSON, or JSON-L (all converted automatically to JSON-L).
Start fine-tuning; training uses LoRA for efficiency.
Step 4: Test and refine
Use the Model Details page to chat with the model in-browser.
Adjust Temperature and Max Tokens to explore how the model behaves.
If outputs are off, refine your dataset and retrain.
Step 5: Deploy or download
Download in Hugging Face format (HF) for use in your own environment.
Or deploy instantly with Minibase Cloud for fast API-based access.
Tips & Best Practices
Use task prefixes: The model supports task-specific prefixes like
summarize:,translate English to French:, orclassify:. Adding these helps ensure accurate outputs.
Keep inputs concise: Because the context window is 512 tokens, overly long inputs may be truncated. Summarize or split long documents.
Match dataset to task: The model performs best when your dataset examples look like your intended real-world use cases.
Quantization options: Models can be downloaded in different fidelity/size trade-offs:
High (Q6_K): production-grade, best quality.
Medium (Q4_K_M): balanced for most general uses.
Low (Q4_K_S): compact, suitable for mobile or edge devices.
Troubleshooting
The model cuts off responses
The model cuts off responses
This typically indicates a mismatch between training data and test input. Solution: Use more examples that resemble the edge-case inputs you plan to use. If you trained with medical examples, don’t test with physics prompts—model performance drops on unfamiliar domains.
This typically indicates a mismatch between training data and test input. Solution: Use more examples that resemble the edge-case inputs you plan to use. If you trained with medical examples, don’t test with physics prompts—model performance drops on unfamiliar domains.
Outputs aren’t accurate for my task
Outputs aren’t accurate for my task
Make sure you’re including the proper task prefix in your instruction (e.g., `summarize:`, `translate English to French:`). Consistent formatting improves results.
Training is slow
Training is slow
Training time grows with dataset size. For large datasets (10k+ examples), expect training to take several hours. Use smaller test datasets for rapid prototyping.