README.md

metadata

license: apache-2.0
tags:
  - merge
  - hindi
  - english
  - Llama2
  - ai4bharat/Airavata
  - BhabhaAI/Gajendra-v0.1
model-index:
  - name: yuj-v1
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 45.65
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 70.1
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 43.78
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 41.69
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 69.85
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 4.78
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=shuvom/yuj-v1
          name: Open LLM Leaderboard

The Model yuj-v1:

The yuj-v1 model is a blend of advanced models strategically crafted to enhance Hindi Language Models (LLMs) effectively and democratically. Its primary goals include catalyzing the development of Hindi and its communities, making significant contributions to linguistic knowledge. The term "yuj," from Sanskrit, signifies fundamental unity, highlighting the integration of sophisticated technologies to improve the language experience for users in the Hindi-speaking community.

Official GGUF version: shuvom/yuj-v1-GGUF

Below are the model which are leverage to build this yuj-v1:

• ai4bharat/Airavata
• BhabhaAI/Gajendra-v0.1

☄️Space to use it (yuj-v1 tryO):

💻 Usage:

First, you need to install some of below packages:

1. Bits and bytes

!pip install bitsandbytes

2. Accelerate (to install the latest version)

!pip install git+https://github.com/huggingface/accelerate.git

3. Usage

# Usage
import torch

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

# load the model in 4-bit quantization
tokenizer = AutoTokenizer.from_pretrained("shuvom/yuj-v1")
model = AutoModelForCausalLM.from_pretrained("shuvom/yuj-v1",torch_dtype=torch.bfloat16,load_in_4bit=True)

prompt = "युज शीर्ष द्विभाषी मॉडल में से एक है"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate
generate_ids = model.generate(inputs.input_ids, max_length=65)
tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

4. Output

युज शीर्ष द्विभाषी मॉडल में से एक है। यह एक उत्पादक मॉडल है जो एक साथ एक ट्रांसफॉर्मर और एक आत्म-ध्यान तंत्रिका नेटवर्क को जोड़ता है। यह एक ट्रांसफॉर्मर वास्तुकला का उपयोग करता है जो एक ट्रांसफॉर्मर मॉडल की तुलना में बहुत अधिक जटिल है।

🧩 Configuration

models:
  - model: sarvamai/OpenHathi-7B-Hi-v0.1-Base
    # no parameters necessary for base model
  - model: ai4bharat/Airavata
    parameters:
      density: 0.5
      weight: 0.5
  - model: BhabhaAI/Gajendra-v0.1
    parameters:
      density: 0.5
      weight: 0.3
merge_method: ties
base_model: sarvamai/OpenHathi-7B-Hi-v0.1-Base
parameters:
  normalize: true
dtype: float16

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	45.97
AI2 Reasoning Challenge (25-Shot)	45.65
HellaSwag (10-Shot)	70.10
MMLU (5-Shot)	43.78
TruthfulQA (0-shot)	41.69
Winogrande (5-shot)	69.85
GSM8k (5-shot)	4.78