📌 Model Overview
Sarus-500M is a 500M-parameter compact causal language model developed by ViuAI. The model is designed for efficient, low-latency reasoning and bilingual text generation across English and Hindi (Hinglish).
Sarus executes an internal Cognitive Reasoning Monologue ([THINK] ... [/THINK]) before producing its final answer, enabling structured step-by-step problem solving in a lightweight architecture.
🌟 Key Features
- 🧠 Cognitive Reasoning: Structured chain-of-thought monologue enclosed in
[THINK] ... [/THINK]tags. - 🇮🇳 Bilingual Fluency: Native support for English, Hindi, and romanized Hinglish.
- ⚡ High Efficiency: Optimized for low-latency edge and server deployment using Grouped Query Attention (GQA).
- 💬 Conversational: Multi-turn dialogue support with standard chat formatting.
📐 Architecture Summary
| Feature | Value |
|---|---|
| Parameters | ~500M |
| Hidden Size | 1280 |
| Layers | 24 |
| Attention Heads | 20 (Query) / 4 (KV) |
| Max Context | 2048 |
| Vocabulary Size | 64,009 |
🚀 Quickstart
Chat Template Format
<|user|>
{user_message}<|endofturn|>
<|assistant|>
[THINK]
{internal_reasoning}
[/THINK]
{final_answer}<|endofturn|>
Python Inference
from transformers import AutoTokenizer, AutoModelForCausalLM
# Load Tokenizer & Model
model_id = "ViuAI/Sarus-500M"
tokenizer = AutoTokenizer.from_pretrained(model_id)
# model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
📄 License
Sarus-500M is licensed under the Apache 2.0 License.
Developed by the ViuAI Team.
- Downloads last month
- 64