FritzStack/Go-QWENemotions-0.6B

Model Description

This is a lightweight model based on Qwen3-0.6B fine-tuned on the GoEmotions dataset for multi-class, multi-label emotion classification.

GoEmotions Dataset

GoEmotions is the largest manually annotated English language fine-grained emotion dataset, consisting of 58,000 Reddit comments labeled with 27 emotion categories or Neutral. The dataset was designed with both psychological validity and data applicability in mind, offering a balanced representation of emotions with 12 positive, 11 negative, and 4 ambiguous emotion categories.

The 27 emotion categories include: admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, and surprise.

Model Architecture

Built on the Qwen3-0.6B architecture, this model provides an efficient solution for emotion detection tasks while maintaining strong performance. The lightweight design (0.6B parameters) makes it suitable for deployment in resource-constrained environments, including edge devices and real-time applications.

Multi-Label Classification

Each input sentence can be assigned multiple emotion labels simultaneously, reflecting the complexity of real emotional expressions where people often experience mixed or overlapping emotions. This multi-label capability is particularly valuable for understanding nuanced emotional content in conversational contexts.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch


tokenizer = AutoTokenizer.from_pretrained(
    "FritzStack/Go-QWENemotions-0.6B", 
    trust_remote_code=True
)

model = AutoModelForCausalLM.from_pretrained(
    "FritzStack/Go-QWENemotions-0.6B",
    load_in_4bit=True,
    device_map="auto",
    trust_remote_code=True
)

def predict_emotions(text, max_new_tokens=50):
    """
    Predict emotions for a given text
    """
    prompt = f"{text}"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.9,
            top_k = 10,
            #repetition_penalty=35.,
            pad_token_id=tokenizer.eos_token_id
        )

    generated_text = tokenizer.decode(
        outputs[0][len(inputs.input_ids[0]):],
        skip_special_tokens=False
    ).strip()

    return generated_text

Uploaded model

• Developed by: FritzStack
• License: apache-2.0
• Finetuned from model : unsloth/qwen3-0.6b-unsloth-bnb-4bit

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.