🌐 Svarix-1-1B-Hi
A Hindi-Optimized Fine-Tuned Variant of Gemma-3-1B-IT
By Pathania Svarix AI — India 🇮🇳
🔷 Overview
Svarix-1-1B-Hi is a fine-tuned model based on google/gemma-3-1b-it, optimized specifically for natural, conversational, Indian-style Hindi.
Where the base Gemma model sounds formal or machine-translated in Hindi, Svarix-1-1B-Hi has been tuned to produce:
- Everyday conversational Hindi
- Semi-formal/office Hindi
- WhatsApp-style replies
- Short, crisp Indian text responses
- Hindi summaries and translations
This release (v1) focuses on pure Hindi optimization.
Support for Hinglish & Roman Hindi is planned for the next update (v1.1).
🧠 Model Description
| Field | Value |
|---|
| Base Model | google/gemma-3-1b-it |
| Architecture | Decoder-only Transformer |
| Size | 1 Billion parameters |
| Format | safetensors |
| Precision | FP16 |
| Optimized For | Hindi conversational tasks |
| Organization | Pathania Svarix Pvt. Ltd. |
🛠️ Training Details
Fine-tuning Objective
The goal was to make Gemma-3-1B sound more:
- Indian
- Context-aware
- Conversational
- Less formal
- More culturally aligned
Dataset Used
- ai4bharat/sangraha (Hindi portion)
A curated Hindi corpus containing diverse textual styles:
- news
- social text
- government docs
- conversational lines
- general Hindi usage
Training Method
- LoRA / QLoRA-based parameter-efficient tuning
- Larger sequence lengths to reduce training time
- Instruction-style formatting matching Gemma chat template
Full fine-tuning logs and training code will be released in future updates.
🎯 Intended Use
This model is ideal for:
- Hindi Chat Assistants
- Customer support bots
- WhatsApp/TG/website chat widgets
- Indian small-scale AI workflows
- Basic Hindi summarization & paraphrasing
- Hindi-English translation (simple level)
- Document reply generation
- Call center or support desk tools
Good for
- < 8 GB GPU setups
- Mobile or edge devices (quantized versions coming soon)
🚫 Not Intended For
- Legal or medical advice
- Financial or compliance-critical answers
- Deep reasoning tasks
- High-risk or safety-critical systems
- Misinformation generation
- Political persuasion
📦 How to Use
Python Example
1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3model = "PathaniaSvarix/Svarix-1-1b-hi"
4
5tok = AutoTokenizer.from_pretrained(model)
6mod = AutoModelForCausalLM.from_pretrained(model)
7
8inp = "एक छोटा सा हिंदी सारांश लिखो: भारत का बजट 2024 क्या कहता है?"
9x = tok(inp, return_tensors="pt")
10
11out = mod.generate(
12 **x,
13 max_new_tokens=200,
14 temperature=0.7
15)
16
17print(tok.decode(out[0], skip_special_tokens=True))
🗣️ Example Prompts
Casual Hindi
Office Hindi
कृपया इस मेल का एक छोटा पेशेवर उत्तर लिखें।Summarization
इस पैराग्राफ का 3 लाइन का सारांश लिखो।Translation
Translate to Hindi: "Please complete the form and submit before 5 pm."
🔄 Roadmap
Planned for v1.1 (Next Release)
- Hinglish optimization
- Roman Hindi training
- Improved chat alignment
- GGUF quantized models for LM Studio / Ollama
- Full training scripts
- Evaluation benchmarks
- HF Space demo (public playground UI)
🛡️ License
This model is a derivative of
google/gemma-3-1b-it
and is therefore distributed under the same Gemma License Terms.
License: Gemma Terms of Use
Full terms:
https://ai.google.dev/gemma/terms
You must comply with the original Gemma license when using or redistributing this model.
🏢 Organization
Developed by:
Pathania Svarix Pvt. Ltd.
Svarix AI Division — India 🇮🇳
Contact (AI Wing):
contact@svarix.ai
⭐ Acknowledgements
Special thanks to:
- Google DeepMind — for the Gemma architecture
- AI4Bharat — for the Sangraha Hindi datasets
- Open-source community
🔔 Notes
This is v1 of the model.
Future versions will significantly improve:
- Hinglish
- Roman Hindi
- General reasoning ability
- Safety alignment
Stay tuned for updates.
