RAS1981/turn-detection-probability-balanced

🇷🇺 Russian Real-Estate Turn Detection (Probability Balanced)

This dataset is designed for training probability-based turn detection models for Russian conversational AI, specifically in the real-estate domain (renting, buying, inquiries).

It follows a contrastive approach: for every complete user query, there is a corresponding incomplete version. This forces the model to learn the subtle semantic and grammatical cues that signal the end of a turn versus a mid-sentence pause.

🎯 Goal

The goal is to fine-tune a Language Model (LLM) to predict whether a user has finished speaking or is just pausing.

• Complete Turn: The model should predict the End-of-Sequence token (<|im_end|>).
• Incomplete Turn: The model should predict the continuation (the next word), assigning a low probability to <|im_end|>.

📊 Dataset Structure

Each example contains a prefix (the conversation history) and a target (what comes next).

Field	Description
prefix	The user's input text (e.g., "Здравствуйте, я по поводу").
target_token	The expected next token. Either a real word (for incomplete) or `<
target_id	The token ID of the target (specific to Qwen tokenizer, but adaptable).
label	complete or incomplete.

Example Pairs

1. Complete Turn (User finished)

• Input: "Здравствуйте, я звоню по объявлению о квартире."
• Target: <|im_end|> (EOS Token)
• Label: complete

2. Incomplete Turn (User paused)

• Input: "Здравствуйте, я звоню по объявлению о..."
• Target: "квартире" (Next word)
• Label: incomplete

⚖️ Balancing & Methodology

• Total Size: ~20,000 examples.
• Distribution: Strictly 50% Complete / 50% Incomplete.
• Source: Synthetic generation using advanced LLMs (Gemini/GPT-4o) simulating realistic phone conversations.
• Filtering: Validated to ensure "incomplete" examples are grammatically cut off and "complete" examples are semantically finished.

🛠️ Usage (Probability-Based Turn Detection)

This dataset is optimized for Single-Token Loss Masking:

1. Feed the prefix into the model.
2. Calculate the loss only on the target token.
3. If label=complete, maximize $P(\text{<|im_end|>})$.
4. If label=incomplete, maximize $P(\text{next_word})$ (which implicitly minimizes $P(\text{<|im_end|>})$).

Inference Logic

At inference time, you only check the probability of <|im_end|>:

probs = softmax(model(input_ids).logits)
eos_prob = probs[EOS_TOKEN_ID]

if eos_prob > threshold (e.g., 0.4):
    return "TURN_COMPLETED"
else:
    return "TURN_INCOMPLETE"