Instructions to use eulogik/Bharat-Tiny-LLM-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use eulogik/Bharat-Tiny-LLM-v2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="eulogik/Bharat-Tiny-LLM-v2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("eulogik/Bharat-Tiny-LLM-v2")
model = AutoModelForCausalLM.from_pretrained("eulogik/Bharat-Tiny-LLM-v2", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use eulogik/Bharat-Tiny-LLM-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "eulogik/Bharat-Tiny-LLM-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eulogik/Bharat-Tiny-LLM-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/eulogik/Bharat-Tiny-LLM-v2

SGLang

How to use eulogik/Bharat-Tiny-LLM-v2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "eulogik/Bharat-Tiny-LLM-v2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eulogik/Bharat-Tiny-LLM-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "eulogik/Bharat-Tiny-LLM-v2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eulogik/Bharat-Tiny-LLM-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use eulogik/Bharat-Tiny-LLM-v2 with Docker Model Runner:
```
docker model run hf.co/eulogik/Bharat-Tiny-LLM-v2
```

Brahmi: Efficient Devanagari Token Injection for Multilingual LLMs (Technical Breakdown)

by GautamKishore - opened 1 day ago

Discussion

GautamKishore

Eulogik org 1 day ago

Brahmi: Efficient Devanagari Token Injection for Multilingual LLMs

Abstract & Problem Overview

Multilingual transformer models built on English-heavy training corpora penalize Indian languages with a severe "Token Tax". Standard Byte-Pair Encoding (BPE) tokenizers (such as Qwen2.5's 151.9k vocabulary) split common Devanagari words into 3 to 4 fragmented byte tokens.

For example, the simple Hindi sentence:

"ज़रूरी बात है क्या करते हो"

Base Qwen2.5 Tokenizer: Burns 26 tokens 🔴
Bharat-Tiny-LLM v2 (Brahmi Tokenizer): Requires only 11 tokens 🟢 (58% Token Savings)

This token bloat inflates memory bandwidth requirements, slows down generation throughput, and forces early context window exhaustion.

Technical Innovation: Brahmi Token Injection

Instead of retraining a 1.5B model from scratch, we developed Brahmi Token Injection — a surgical two-stage vocabulary expansion and embedding alignment technique:

Vocabulary Expansion: Identified top 300 recurring Devanagari subwords from a 50GB Hindi corpus and injected them into the tokenizer dictionary (expanding vocabulary from 151,936 to 152,236 tokens).
Stage 1 (Embedding Alignment): Initialized new embedding matrix rows. Froze the transformer decoder backbone and trained strictly the 300 new token embeddings to align their representations with existing semantic space.
Stage 2 (LoRA Fine-Tuning): Attached rank-16 PEFT LoRA adapters ($lpha=32$) across attention projection matrices (q_proj, k_proj, v_proj, o_proj) for downstream instruction tuning.

Empirical Benchmarks & Quantitative Results

Metric	Base Qwen2.5-1.5B	Bharat-Tiny-LLM v2	Technical Advantage
Tokens for 1,000 Hindi Chars	~950 tokens	~630 tokens	33.8% Token Reduction
Hindi Inference Speed	50 tok/s	68 tok/s	+36% Faster Throughput
Validation Loss (Hindi Corpus)	2.776	1.837	52.5% Loss Reduction (Perplexity: 16.1 → 6.3)
Model Size (Q4 Affine)	880 MB	880 MB	Runs in <3.8 GB RAM (Apple Silicon / PyTorch)

Code Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "eulogik/Bharat-Tiny-LLM-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")

prompt = "भारत की सांस्कृतिक विविधता के बारे में बताइए:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Open Weights & Apache 2.0 License. Built by Eulogik (@GautamKishore ).

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment