Understanding the different training tasks supported by ModelForge.

ModelForge supports three primary training tasks, each optimized for different use cases. The task you choose determines how your model processes data and what kind of outputs it generates.

Task ID: text-generation

Generate text based on input prompts. This is the most versatile task type, suitable for:

✅ Chatbots and conversational AI
✅ Instruction following
✅ Code generation
✅ Creative writing
✅ Story completion
✅ Dialog systems

{"input": "What is machine learning?", "output": "Machine learning is a subset of AI..."}
{"input": "Write a Python function to sort a list", "output": "def sort_list(lst):\n    return sorted(lst)"}
{"input": "Continue this story: Once upon a time", "output": "there was a brave knight who..."}

{
  "task": "text-generation",
  "model_name": "meta-llama/Llama-3.1-8B-Instruct",
  "dataset": "/path/to/dataset.jsonl",
  "provider": "unsloth",
  "strategy": "qlora",
  "num_train_epochs": 3
}

Task ID: summarization

Condense long text into shorter, coherent summaries. Ideal for:

✅ Document summarization
✅ Article condensing
✅ Meeting notes generation
✅ Email summarization
✅ Report abstracts

{"input": "Long article or document text here...", "output": "Brief summary of the content."}
{"input": "Detailed meeting transcript...", "output": "Key points and action items."}

Input: Full-length text (can be several paragraphs)
Output: Concise summary (typically 10-30% of original length)

{
  "task": "summarization",
  "model_name": "google/flan-t5-large",
  "dataset": "/path/to/summaries.jsonl",
  "provider": "huggingface",
  "strategy": "sft",
  "max_seq_length": 2048,
  "num_train_epochs": 3
}

Task ID: extractive-question-answering

Extract answers from text passages. Perfect for:

✅ RAG (Retrieval-Augmented Generation) systems
✅ Document search and retrieval
✅ FAQ bots
✅ Information extraction
✅ Reading comprehension

{"context": "Python is a programming language created by Guido van Rossum.", "question": "Who created Python?", "answer": "Guido van Rossum"}
{"context": "The capital of France is Paris, located on the Seine River.", "question": "What is the capital of France?", "answer": "Paris"}

Fields:

• context: The passage containing the answer
• question: The question to be answered
• answer: The exact answer text (must appear in context)

{
  "task": "extractive-question-answering",
  "model_name": "deepset/roberta-base-squad2",
  "dataset": "/path/to/qa-dataset.jsonl",
  "provider": "huggingface",
  "strategy": "sft",
  "num_train_epochs": 3
}

Need a chatbot or instruction follower?
    → text-generation

Need to condense documents?
    → summarization

Need to extract specific information from documents?
    → extractive-question-answering

{
  "task": "text-generation",
  "max_seq_length": 2048,  // Longer for conversations
  "learning_rate": 2e-4,
  "num_train_epochs": 3
}

{
  "task": "summarization",
  "max_seq_length": 1024,  // Shorter is often sufficient
  "learning_rate": 3e-4,   // Can use higher LR
  "num_train_epochs": 5    // May need more epochs
}

{
  "task": "extractive-question-answering",
  "max_seq_length": 512,   // Typically shorter contexts
  "learning_rate": 3e-5,   // Lower LR for precision
  "num_train_epochs": 3
}

1. Quality over quantity: 100 high-quality examples > 1000 poor examples
2. Diverse examples: Cover various scenarios your model will encounter
3. Balanced data: Include positive and negative examples
4. Clean formatting: Use proper JSON/JSONL format
5. Validation split: Reserve 10-20% for evaluation

ModelForge automatically evaluates your models using task-specific metrics:

• Perplexity: Lower is better (measures prediction confidence)
• Loss: Training and validation loss

• ROUGE scores: ROUGE-1, ROUGE-2, ROUGE-L
• Loss: Sequence-to-sequence loss

• Exact Match (EM): Percentage of exact answer matches
• F1 Score: Token-level F1 between prediction and ground truth
• Loss: Classification/extraction loss

Problem: Using text-generation for summarization
Solution: Use the summarization task for better results and faster training

Problem: Using {"text": "...", "label": "..."} for text-generation
Solution: Use {"input": "...", "output": "..."} format

Problem: Using BERT for text-generation
Solution: Use decoder-only models (Llama, Mistral) for generation

• Dataset Formats - Detailed dataset preparation guide
• Configuration Guide - All configuration options
• Hardware Profiles - Optimize for your GPU
• Model Recommendations - Choose the right model

Choose the right task for your use case! Each task is optimized for specific scenarios.