vllm-sr/mmbert-feedback-detector-merged

mmBERT Feedback Detector (Merged)

A multilingual 4-class user feedback classifier built on jhu-clsp/mmBERT-base. This model classifies user responses into satisfaction categories to help understand user intent in conversational AI systems.

Model Description

This is the merged model (LoRA weights merged into base model) for direct inference without PEFT. For the LoRA adapter version, see vllm-sr/mmbert-feedback-detector-lora.

Labels

Label	ID	Description
SAT	0	User is satisfied with the response
NEED_CLARIFICATION	1	User needs more explanation or clarification
WRONG_ANSWER	2	User indicates the response is incorrect
WANT_DIFFERENT	3	User wants alternative options or different response

Performance

Metric	Score
Accuracy	96.89%
F1 Macro	96.88%
F1 Weighted	96.88%

Per-Class Performance

Class	F1 Score
SAT	100.0%
NEED_CLARIFICATION	99.7%
WRONG_ANSWER	94.0%
WANT_DIFFERENT	93.8%

Multilingual Support

Thanks to mmBERT's multilingual pretraining (256k vocabulary, 100+ languages), this model achieves excellent cross-lingual transfer:

Language	Accuracy
🇺🇸 English	100%
🇪🇸 Spanish	100%
🇫🇷 French	100%
🇩🇪 German	100%
🇨🇳 Chinese	100%
🇯🇵 Japanese	100%
🇰🇷 Korean	100%
🇸🇦 Arabic	100%
🇵🇹 Portuguese	100%
🇷🇺 Russian	100%

Usage

With Transformers

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_name = "vllm-sr/mmbert-feedback-detector-merged"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example: Classify user feedback
text = "Thanks, that's exactly what I needed!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    pred = probs.argmax().item()

labels = ["SAT", "NEED_CLARIFICATION", "WRONG_ANSWER", "WANT_DIFFERENT"]
print(f"Prediction: {labels[pred]} ({probs[0][pred]:.1%})")
# Output: Prediction: SAT (100.0%)

With Pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="vllm-sr/mmbert-feedback-detector-merged"
)

# English
result = classifier("Thanks, that's helpful!")
print(result)  # [{'label': 'SAT', 'score': 0.999...}]

# Spanish (cross-lingual transfer)
result = classifier("¡Gracias, eso es muy útil!")
print(result)  # [{'label': 'SAT', 'score': 0.999...}]

# Chinese
result = classifier("谢谢，这很有帮助！")
print(result)  # [{'label': 'SAT', 'score': 0.98...}]

Training Details

• Base Model: jhu-clsp/mmBERT-base
• Method: LoRA fine-tuning + merge
• LoRA Rank: 16
• LoRA Alpha: 32
• Learning Rate: 2e-5
• Batch Size: 32
• Epochs: 5
• Max Length: 512
• Dataset: vllm-sr/feedback-detector-dataset

Use Cases

• Conversational AI: Understand if users are satisfied with chatbot responses
• Customer Support: Route dissatisfied users to human agents
• Quality Monitoring: Track response quality across languages
• Feedback Analysis: Categorize user feedback automatically

Related Models

• vllm-sr/mmbert-feedback-detector-lora - LoRA adapter version
• vllm-sr/mmbert-intent-classifier-merged - Intent classification
• vllm-sr/mmbert-fact-check-merged - Fact checking
• vllm-sr/mmbert-jailbreak-detector-merged - Security

Citation

@misc{mmbert-feedback-detector,
  title={mmBERT Feedback Detector},
  author={vLLM Semantic Router Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/vllm-sr/mmbert-feedback-detector-merged}
}

License

Apache 2.0