eulogik/Bharat-Tiny-LLM-v2
Bharat-Tiny-LLM v2 by eulogik is a 1.5 billion parameter language model built on Qwen2.5-1.5B, specifically optimized for Hindi and Hinglish text generation. It utilizes "Brahmi Token Injection" to integrate 300 Devanagari subword tokens, achieving 33.8% token compression and 36% faster Hindi inference compared to its base model. This model is ideal for applications requiring efficient and accurate processing of Indian languages, particularly Hindi and Hinglish.
Loading preview...
Bharat-Tiny-LLM v2: Optimized for Hindi and Hinglish
Bharat-Tiny-LLM v2 is a 1.5 billion parameter open-weights language model developed by eulogik, based on the Qwen2.5-1.5B architecture. Its core innovation is Brahmi Token Injection, a technique that expands the tokenizer's vocabulary with 300 Devanagari subword tokens. This method significantly reduces the "Token Tax" typically imposed by English-centric tokenizers on Indian scripts.
Key Capabilities & Advantages
- Enhanced Hindi Efficiency: Achieves 33.8% fewer tokens for Hindi text (up to 58% on chat prompts) and a 36% speed boost in Hindi inference throughput.
- Improved Language Understanding: Demonstrates a 52.5% reduction in validation loss on Hindi corpora, leading to better perplexity (16.1 down to 6.3).
- Optimized Architecture: The training methodology involves a two-stage process: initial embedding alignment for the new Brahmi tokens, followed by LoRA fine-tuning on attention projections.
- Broad Compatibility: Provided as unquantized PyTorch / HuggingFace Transformers weights, ensuring compatibility with CUDA GPUs, vLLM, TGI, and Google Colab.
When to Use This Model
Bharat-Tiny-LLM v2 is particularly well-suited for:
- Applications requiring efficient and fast text generation in Hindi and Hinglish.
- Scenarios where token compression is critical for cost or performance, especially with Indian languages.
- Developers looking for a compact yet powerful LLM for Indic language processing tasks.
For Apple Silicon users, an MLX quantized version is available at eulogik/Bharat-Tiny-LLM-v2-MLX, and a GGUF version for llama.cpp is at eulogik/Bharat-Tiny-LLM-GGUF.