> ## Documentation Index > Fetch the complete documentation index at: https://docs.monostate.ai/llms.txt > Use this file to discover all available pages before exploring further. # Training Tasks > What you can train models to do # Types of Training Tasks AI Training supports many different tasks. Each task is optimized for specific types of problems you want to solve. ## Text Tasks ### Text Classification **What it does**: Sorts text into categories you define. **Real-world examples**: * Email spam detection * Customer feedback sentiment (happy/unhappy) * Support ticket routing * Content moderation * Language detection **What you need**: * Text samples * Category labels for each sample * At least 50 examples per category (more is better) **Example data format**: ```csv theme={null} text,label "Great product, highly recommend!",positive "Terrible experience, waste of money",negative "The item arrived damaged",negative "Excellent customer service",positive ``` ### Token Classification (NER) **What it does**: Labels specific words or phrases in text. **Real-world examples**: * Extract names, dates, locations from documents * Identify product mentions in reviews * Find medical terms in clinical notes * Highlight important contract clauses * Tag parts of speech **What you need**: * Text with marked entities * BIO format (Beginning, Inside, Outside) labels * Hundreds of annotated examples **Example data format**: ``` John B-PERSON Smith I-PERSON visited O New B-LOCATION York I-LOCATION yesterday B-DATE ``` ### Sequence to Sequence **What it does**: Transforms one text into another. **Real-world examples**: * Language translation * Text summarization * Question answering * Text correction * Paraphrasing **What you need**: * Input text * Desired output text * Pairs of input-output examples **Example data format**: ```csv theme={null} input_text,target_text "The quick brown fox jumps over the lazy dog","Le renard brun rapide saute par-dessus le chien paresseux" "Hello, how are you?","Bonjour, comment allez-vous?" ``` ### Text Generation (LLM Fine-tuning) **What it does**: Teaches language models new knowledge or behaviors. **Real-world examples**: * Custom chatbots * Domain-specific assistants * Code generation * Creative writing * Technical documentation **What you need**: * Conversation examples or instruction-response pairs * Optional: Preference data for RLHF * Can work with as few as 100 examples **Example conversation format**: ```json theme={null} { "messages": [ {"role": "user", "content": "What is photosynthesis?"}, {"role": "assistant", "content": "Photosynthesis is the process plants use to convert sunlight into energy..."} ] } ``` ## Image Tasks ### Image Classification **What it does**: Identifies what's in an image. **Real-world examples**: * Product quality inspection * Medical image diagnosis * Wildlife identification * Document type classification * Facial expression recognition **What you need**: * Images (JPG, PNG) * Category label for each image * At least 100 images per category **Folder structure**: ``` data/ cats/ cat1.jpg cat2.jpg dogs/ dog1.jpg dog2.jpg ``` ### Object Detection **What it does**: Finds and locates multiple objects in images. **Real-world examples**: * Inventory counting * Security monitoring * Autonomous driving * Defect detection * People counting **What you need**: * Images with bounding boxes * Labels for each box * COCO or YOLO format annotations ## Structured Data Tasks ### Tabular Classification **What it does**: Predicts categories from spreadsheet-like data. **Real-world examples**: * Customer churn prediction * Fraud detection * Disease diagnosis * Credit approval * Equipment failure prediction **What you need**: * CSV with features and labels * Numerical and categorical columns * Clean, preprocessed data **Example data**: ```csv theme={null} age,income,credit_score,approved 25,45000,720,yes 35,65000,650,no 42,85000,780,yes ``` ### Tabular Regression **What it does**: Predicts continuous values from structured data. **Real-world examples**: * House price prediction * Sales forecasting * Energy consumption estimation * Stock price prediction * Delivery time estimation **What you need**: * CSV with features and target values * Numerical target column * Historical data ## Advanced Training Methods ### Supervised Fine-Tuning (SFT) Standard training with examples and correct answers. **Use when**: You have good quality labeled data. ### DPO (Direct Preference Optimization) Train models using preference comparisons. **Use when**: You have examples of good vs bad outputs. **Data format**: ```json theme={null} { "prompt": "Write a greeting", "chosen": "Hello! How can I help you today?", "rejected": "sup" } ``` ### ORPO (Odds Ratio Preference Optimization) Similar to DPO but more stable training. **Use when**: DPO training is unstable or overfitting. ### Reward Modeling Train a model to score outputs. **Use when**: Building a reward model for RLHF. ### PPO (Proximal Policy Optimization) Reinforcement learning from feedback. **Use when**: You have a reward model and want to optimize against it. ## Task Selection Guide ### Based on Your Data | If you have... | Choose this task | | ----------------------- | --------------------------------- | | Text + categories | Text Classification | | Text with entity labels | Token Classification | | Input/output text pairs | Sequence to Sequence | | Conversations | LLM Fine-tuning | | Images + labels | Image Classification | | Spreadsheet data | Tabular Classification/Regression | ### Based on Your Goal | If you want to... | Choose this task | | ------------------------ | -------------------- | | Sort things into buckets | Classification | | Extract information | Token Classification | | Transform text | Sequence to Sequence | | Create a chatbot | LLM Fine-tuning | | Predict numbers | Regression | | Find objects | Object Detection | ### Based on Difficulty **Easiest to start**: 1. Text Classification 2. Image Classification 3. Tabular Classification **Medium difficulty**: 1. Token Classification 2. Sequence to Sequence 3. LLM Fine-tuning (SFT) **Advanced**: 1. DPO/ORPO training 2. Object Detection 3. PPO/RLHF ## Data Requirements ### Minimum Data Needed | Task | Absolute Minimum | Good Starting Point | Production Quality | | -------------------- | ---------------- | ------------------- | ------------------ | | Text Classification | 50 per class | 500 per class | 5,000+ per class | | Token Classification | 100 documents | 1,000 documents | 10,000+ documents | | Seq2Seq | 100 pairs | 1,000 pairs | 10,000+ pairs | | LLM Fine-tuning | 50 examples | 500 examples | 5,000+ examples | | Image Classification | 100 per class | 1,000 per class | 10,000+ per class | | Tabular | 500 rows | 5,000 rows | 50,000+ rows | ### Data Quality Matters Better to have 100 high-quality examples than 1,000 poor ones: * Accurate labels * Consistent formatting * Representative of real-world use * Balanced across categories ## Multi-Task Training You can train models for multiple tasks simultaneously: ### Benefits * Share knowledge between tasks * More efficient use of data * Single model deployment ### Example Train one model to: * Classify sentiment * Extract entities * Summarize text All with the same base model. ## Task-Specific Settings ### Text Tasks * **Max sequence length**: How much text to process * **Tokenizer**: How to split text into tokens * **Special tokens**: Task-specific markers ### Image Tasks * **Image size**: Resolution to use * **Augmentation**: Rotation, flip, crop * **Normalization**: Pixel value scaling ### Tabular Tasks * **Feature engineering**: Creating new columns * **Scaling**: Normalizing numeric values * **Encoding**: Handling categorical variables ## Evaluation Metrics Different tasks use different metrics: | Task | Common Metrics | | -------------------- | ------------------------------- | | Classification | Accuracy, F1, Precision, Recall | | Token Classification | Entity-level F1, Token accuracy | | Seq2Seq | BLEU, ROUGE, BERTScore | | Generation | Perplexity, Human evaluation | | Regression | MSE, MAE, R² | | Object Detection | mAP, IoU | ## Combining Tasks ### Pipeline Approach Chain tasks together: 1. Classification → Route to specialized model 2. NER → Extract entities → Generate response 3. Translate → Summarize → Classify sentiment ### Multi-Modal Tasks Combine different data types: * Image + Text → Visual QA * Audio + Text → Speech recognition * Video + Text → Video understanding ## Next Steps Ready to dive deeper? Choose the right model architecture Train your first model