---
language: en
license: mit
model_name: gemma-3b-chat-support
tags:
- chat-support
- customer-service
- gemma-3b
- fine-tuned
- schema-guided-dialogue
base_model:
- google/gemma-3-1b-it
metrics:
- bleu
---

<h1 style="text-align: center;">Gemma 3B Chat Support Assistant</h1>
<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/65c485b1e5afa5f1a71fd1eb/8qX0yxQDXZ1gG2NV3MypD.png" alt="Image" width="650" height="400">
</p>

<!-- A fine-tuned version of Google's Gemma 3B model optimized for customer support conversations across multiple domains, capable of handling multi-turn dialogues with context retention. -->

## Model Details

This model is a fine-tuned version of [google/gemma-3-1b-it](https://huggingface.co/google/gemma-3-1b-it).
It has been trained using [TRL](https://github.com/huggingface/trl). Specifically optimized for customer support conversations across multiple domains. It leverages the Schema-Guided Dialogue (SGD) dataset to provide helpful, contextual responses.

## Capabilities

This model can help with:

* **Conversational Support** - Handles multi-turn dialogues with context retention across various customer inquiries
* **Information Retrieval** - Provides relevant information based on user requests across multiple domains
* **Context Management** - Maintains conversation history to provide coherent and contextually appropriate responses
* **Multi-domain Assistance** - Flexibly switches between different domains and topics as the conversation evolves
* **User Engagement** - Creates dynamic, personalized responses that adapt to the changing context of a conversation

## Quick start

```python
from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="iprajwaal/gemma-3b-chat-support", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

```

## Advanced Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "iprajwaal/gemma-3b-chat-support"
model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

messages = [
{"role": "system", "content": "You are a helpful customer support assistant."},
{"role": "user", "content": "I need to book a hotel in New York for next weekend."}
]

response = pipe(messages, max_new_tokens=100)
print(response[0]["generated_text"])
```

## Model Details

* **Base Model**: google/gemma-3-1b-it
* **Fine-tuning Method**: LoRA with 4 epochs
* **Architecture**: Gemma 3B (instruction-tuned variant)
* **Language**: English
* **License**: MIT

## Training Data

This model was fine-tuned on a carefully filtered subset of the [Schema-Guided Dialogue dataset](https://github.com/google-research-datasets/dstc8-schema-guided-dialogue) (SGD). The SGD dataset consists of over 20,000 annotated multi-domain, task-oriented conversations between users and virtual assistants spanning 20 domains, including banking, events, media, calendar, travel, weather, and accommodation.

For fine-tuning purposes, we extracted only the user messages and assistant responses, filtering out the complex annotations while preserving the natural conversational flow. This approach allowed us to create a clean dataset focused on high-quality customer support interactions.

## Intended Uses

This model is designed for:

* Customer support automation
* Virtual assistants
* Chat interfaces
* Conversational AI applications
* Multi-domain dialogue systems

## Limitations

* May not have knowledge of events after its training data cutoff (May 2025)
* Performance may vary on domains not well-represented in the training data
* The model may occasionally generate incorrect information when uncertain
* As with all language models, outputs should be reviewed for accuracy in critical applications

### Framework versions

* TRL: 0.15.2
* Transformers: 4.51.3
* Pytorch: 2.6.0+cu124
* Datasets: [Schema-Guided Dialogue dataset](https://github.com/google-research-datasets/dstc8-schema-guided-dialogue) (SGD)
* Tokenizers: 0.21.1

## Citations

Cite TRL as:

```bibtex
@misc{iprajwaal,
    title        = {{Gemma-3b-chat-support: A Fine-tuned Customer Support Assistant}},
    author       = {Prajwal Kumbar},
    year         = 2025,
    note         = {Fine-tuned using the Schema-Guided Dialogue dataset and the TRL library},
    publisher    = {Hugging Face},
    howpublished = {\url{https://huggingface.co/iprajwaal/gemma-3b-chat-support}}
}
```