Building Chat Models

1# Chat model architecture components
2from langtrain.chat import (
3    ConversationManager,
4    ResponseGenerator, 
5    SafetyFilter,
6    ChatModel
7)
8
9# Initialize chat components
10conversation_mgr = ConversationManager(
11    max_history=10,          # Keep last 10 exchanges
12    context_window=2048,     # Token limit for context
13    memory_type="sliding"    # How to handle long conversations
14)
15
16response_gen = ResponseGenerator(
17    model_name="microsoft/DialoGPT-large",
18    temperature=0.7,         # Response creativity
19    max_length=150,          # Response length limit
20    repetition_penalty=1.2   # Avoid repetitive responses
21)
22
23safety_filter = SafetyFilter(
24    filter_toxic=True,       # Block toxic content
25    filter_personal=True,    # Remove personal info
26    custom_filters=[]        # Add custom filters
27)
28
29# Combine into chat model
30chat_model = ChatModel(
31    conversation_manager=conversation_mgr,
32    response_generator=response_gen,
33    safety_filter=safety_filter
34)

1from langtrain.data import ConversationDataset, ConversationProcessor
2
3# Example conversation data format
4conversations = [
5    {
6        "id": "conv_001",
7        "messages": [
8            {"role": "user", "content": "Hello! How are you today?"},
9            {"role": "assistant", "content": "Hello! I'm doing well, thank you for asking. How can I help you today?"},
10            {"role": "user", "content": "I need help with my Python code"},
11            {"role": "assistant", "content": "I'd be happy to help with your Python code! Please share what you're working on."}
12        ],
13        "metadata": {
14            "domain": "programming_help",
15            "quality_score": 8.5,
16            "language": "en"
17        }
18    }
19]
20
21# Process conversation data
22processor = ConversationProcessor(
23    max_turns=20,           # Maximum turns per conversation
24    min_turn_length=5,      # Minimum tokens per turn
25    quality_threshold=7.0,   # Quality filtering
26    balance_speakers=True    # Ensure balanced conversations
27)
28
29# Create dataset
30dataset = ConversationDataset.from_conversations(
31    conversations,
32    processor=processor,
33    train_split=0.8,
34    validation_split=0.1,
35    test_split=0.1
36)
37
38# Data augmentation for better training
39augmented_dataset = processor.augment_conversations(
40    dataset,
41    techniques=[
42        "paraphrase_user_inputs",    # Rephrase user messages
43        "add_conversation_starters",  # Generate opening messages
44        "expand_short_responses",     # Elaborate brief responses
45        "inject_persona_consistency"  # Maintain character consistency
46    ]
47)
48
49print(f"Original dataset: {len(dataset)} conversations")
50print(f"Augmented dataset: {len(augmented_dataset)} conversations")

1from langtrain import ConversationalTrainer
2from langtrain.models import AutoModelForCausalLM
3from transformers import AutoTokenizer
4
5# Initialize conversational trainer
6trainer = ConversationalTrainer(
7    model_name="microsoft/DialoGPT-medium",
8    
9    # Conversation-specific settings
10    conversation_template="chatml",  # or "alpaca", "vicuna"
11    max_conversation_length=2048,
12    response_max_length=256,
13    
14    # Training parameters
15    num_epochs=3,
16    batch_size=4,
17    gradient_accumulation_steps=4,
18    learning_rate=5e-5,
19    warmup_ratio=0.1,
20    
21    # Conversation-aware training
22    turn_based_loss=True,       # Calculate loss per turn
23    context_aware_training=True, # Use full conversation context
24    response_quality_weighting=True, # Weight high-quality responses more
25    
26    # Output settings
27    output_dir="./models/my-chatbot",
28    save_steps=500,
29    eval_steps=250,
30    logging_steps=50
31)
32
33# Prepare conversation data for training
34train_dataset = trainer.prepare_conversation_data(
35    conversations=augmented_dataset["train"],
36    tokenizer_name="microsoft/DialoGPT-medium",
37    add_special_tokens=True,
38    padding_strategy="longest"
39)
40
41eval_dataset = trainer.prepare_conversation_data(
42    conversations=augmented_dataset["validation"],
43    tokenizer_name="microsoft/DialoGPT-medium"
44)
45
46# Custom conversation metrics
47def compute_conversation_metrics(eval_pred):
48    predictions, labels = eval_pred
49    
50    # Calculate conversation-specific metrics
51    bleu_score = trainer.calculate_bleu(predictions, labels)
52    rouge_scores = trainer.calculate_rouge(predictions, labels)
53    perplexity = trainer.calculate_perplexity(predictions, labels)
54    coherence_score = trainer.calculate_coherence(predictions, labels)
55    
56    return {
57        "bleu": bleu_score,
58        "rouge_l": rouge_scores["rouge_l"],
59        "perplexity": perplexity,
60        "coherence": coherence_score
61    }
62
63trainer.compute_metrics = compute_conversation_metrics
64
65# Start conversational fine-tuning
66trainer.train(
67    train_dataset=train_dataset,
68    eval_dataset=eval_dataset
69)
70
71# Save the conversational model
72trainer.save_model()
73trainer.save_conversation_config()