Malaysian-Turn-Detector-Qwen3-0.6B

Fine-tuned Qwen3-0.6B for real-time turn-end detection in Malaysian multilingual call center conversations.

The model predicts P(<|im_end|>) — the probability that a speaker has finished their turn. Designed for low-latency voice agent pipelines (e.g. LiveKit) to determine when to respond.

How It Works

Given a conversation so far, the model outputs the probability of <|im_end|> as the next token:

• P(im_end) > 0.5 → speaker is done talking (turn complete)
• P(im_end) < 0.5 → speaker is still talking (turn incomplete)

Usage

import torch
import math
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Scicom-intl/Malaysian-Turn-Detector-Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16).cuda().eval()

IM_END_ID = tokenizer.convert_tokens_to_ids("<|im_end|>")

def get_turn_end_prob(text):
    """Returns probability that the speaker's turn is complete."""
    # Strip trailing <|im_end|> so the model predicts whether to emit it
    if text.endswith("<|im_end|>"):
        text = text[:-len("<|im_end|>")]
    inputs = tokenizer(text, return_tensors="pt").to("cuda")
    with torch.no_grad():
        logits = model(**inputs).logits
    prob = F.softmax(logits[0, -1], dim=-1)[IM_END_ID].item()
    return prob

# Complete turn - should be high probability
text = "<|im_start|>user\nHello, saya nak tanya pasal bil saya.<|im_end|>\n<|im_start|>assistant\nBoleh, sila berikan nombor akaun anda."
prob = get_turn_end_prob(text)
print(f"P(turn complete) = {prob:.4f}")  # ~0.74

# Incomplete turn - should be low probability
text = "<|im_start|>user\nHello, saya nak tanya pasal bil saya.<|im_end|>\n<|im_start|>assistant\nBoleh, sila berikan nombor"
prob = get_turn_end_prob(text)
print(f"P(turn complete) = {prob:.4f}")  # ~0.00

Eval Results

Test set: 1200 samples (600 positive + 600 negative), 50 conversations per language pair.

Overall (threshold = 0.5)

Per Language

Threshold Sweep

Probability Distribution

Training

• Base model: Qwen/Qwen3-0.6B
• Training data: Positive samples only (complete conversations ending with <|im_end|>)
• Loss: Liger Fused Linear Cross Entropy
• Attention: Flash Attention 2
• Precision: bfloat16
• Block size: 8192 (multipacked)
• Batch size: 4 x 8 gradient accumulation
• Learning rate: 2e-5 (constant)
• Epochs: 1

Training Data Sources

WandB

• v1 training run
• v2 training run

Source code

Source code at https://github.com/Scicom-AI-Enterprise-Organization/small-ablation/tree/main/turn-detector