Gonyai-v1 is a 160M parameter Large Language Model (LLM) built from scratch to capture the poetic essence and cultural depth of the Konkani language (Goan dialect).
Unlike models fine-tuned from generic multilingual bases, Gonyai-v1 uses a custom architecture optimized specifically for Devanagari script nuances and Konkani linguistic patterns.
- Custom Architecture: Built on
KonkanGPT, featuring Rotary Positional Embeddings (RoPE), RMSNorm, and SwiGLU activation functions. - Pure Konkani Tokenizer: A custom 32k Byte-Level BPE tokenizer trained exclusively on Konkani corpora for high semantic density.
- Curated Training Data: Trained on the Konkani Books Corpus-v2, comprising literature, poetry, and regional news.
- Efficiency: At 160M parameters, it is optimized for low-latency inference and edge deployment.
pip install transformers torch accelerateGonyai-v1 was tested against sub-1B global models to evaluate its efficiency in handling the Konkani language.
| Model | Parameters | Token Efficiency (Lower = Native) | Speed (Tokens/Sec) |
|---|---|---|---|
| Gonyai-v1 | 160M | 5.00 | 65.96 |
| Qwen2.5-0.5B | 500M | 6.57 | 33.27 |
| SmolLM2-360M | 360M | 7.85 | 27.00 |
- Efficiency: Gonyai-v1 is ~35% more efficient at representing Konkani text than Qwen2.5 due to its native tokenizer.
- Latency: It delivers 2x higher throughput than larger models, making it ideal for edge deployment.
- Limitations: As a 160M base model, it focuses on linguistic fluency over world knowledge. It may hallucinate facts or struggle with complex logic.
For the best poetic and coherent results, use the following configuration. This prevents the model from "drifting" into repetitive or unrelated topics.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "omdeep22/Gonyai-v1"
device = "cuda" if torch.cuda.is_available() else "cpu"
# 1. Load Tokenizer and Model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
trust_remote_code=True,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
).to(device)
# 2. Prepare Prompt (Base Model Format)
prompt = "गोंयच्या पावसाचेर एक कविता बरोव."
full_text = f"<|user|>\n{prompt}\n<|assistant|>\n"
# add_token_type_ids=False is critical for this custom architecture
inputs = tokenizer(full_text, return_tensors="pt", add_token_type_ids=False).to(device)
# 3. Optimized Generation
with torch.inference_mode():
outputs = model.generate(
**inputs,
max_new_tokens=100,
temperature=0.4, # Balanced for creativity
repetition_penalty=1.2, # Prevents loops in small models
do_sample=True,
pad_token_id=tokenizer.eos_token_id
)
# 4. Decode
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()
print(f"Assistant: {response}")| Component | Specification |
|---|---|
| Total Parameters | 160 Million |
| Hidden Layers | 12 |
| Attention Heads | 12 |
| Embedding Dim | 768 |
| Context Window | 2048 Tokens |
| Vocabulary Size | 32,000 |
Gonyai-v1 is evolving. The current roadmap focuses on high-concurrency deployment and architectural scaling:
- Backend Scaling: Developing a pipeline using AWS ECS Fargate and Amazon SQS to handle up to 1M concurrent users.
- Quantization: Exporting to GGUF/ONNX formats for faster CPU-based inference on mobile devices.
- Global Cache: Implementing Redis caching to reduce redundant compute for common Konkani queries.
- UI/UX: A dedicated Streamlit-based web interface for community testing.
Contributions are what make the open-source community an amazing place to learn, inspire, and create.
- Fork the Project.
- Create your Feature Branch (
git checkout -b feature/NewFeature). - Commit your Changes (
git commit -m 'Add some NewFeature'). - Push to the Branch (
git push origin feature/NewFeature). - Open a Pull Request.
Distributed under the MIT License. See LICENSE for more information.
- Author: Omdeep (GitHub: @Omdeepb69)
- Model Page: omdeep22/Gonyai-v1
- Dataset: omdeep22/Konkani_books_corpus-v2
Built with ❤️ for the Konkani community.