Hugging Face's logo

krutrim-ai-labs
/
Krutrim-2-instruct

PyTorch
mistral
Krutrim
language-model
Model card Files Community
9
Krutrim-2-instruct / README.md
krutrim-admin's picture
krutrim-admin
Updated gpt-4o scores
6b30258 verified
preview code
|
raw
history blame contribute delete
17.6 kB
metadata
language:
  - en
  - hi
  - bn
  - mr
  - te
  - ta
  - kn
  - ml
  - gu
  - as
  - pa
tags:
  - Krutrim
  - language-model
license: other
license_name: krutrim-community-license-agreement-version-1.0
license_link: LICENSE.md

Krutrim-2

Static Badge Static Badge Static Badge Static Badge

Model Overview

Krutrim-2 is a 12B parameter language model developed by the OLA Krutrim team. It is built on the Mistral-NeMo 12B architecture and trained across various domains, including web data, code, math, Indic languages, Indian context data, synthetic data, and books. Following pretraining, the model was finetuned for instruction following on diverse data covering a wide range of tasks, including knowledge recall, math, reasoning, coding, safety, and creative writing.

After fine-tuning, the model underwent Direct Preference Optimization (DPO) to enhance alignment across multiple aspects. DPO was applied to improve response helpfulness, safety, and reasoning.

The model delivers best-in-class performance across Indic tasks and a promising performance on English benchmarks equivalent to models 5-10x the size. We present details of the model architecture, pre-training, post-training and evaluation results. We also publicly release the post-trained versions of the model. We are continuously improving the model through post-training techniques such as RLHF.

Krutrim 2

Key Features

  • 12B parameter dense transformer model leading to better generalization compared to Krutrim-1 7B;
  • Supports context up to 128K tokens making it suitable for long multi-turn conversations, long-form generations, document translations and others;
  • Delivers competitive performance on most English benchmarks and HumanEval coding task;
  • Natively multilingual delivering best-in-class performance on Indic benchmarks;
  • Matches or exceeds performance of models much larger (5-10x) on multilingual Indic generation tasks including creative writing, summarization, and translation;
  • Stronger Indian cultural context relevance - scored the highest in manual evaluation with multiple models in an anonymised setting;
  • Delivers top-3 performance on 5 (out of 7) tasks in BharatBench among much larger open source and commercial models.
  • Available in instruction-tuned version

Model Developer

  • OLA Krutrim Team

Model Dates

  • Krutrim-2 was trained between Dec 2024 and Jan 2025.

Release History

Model Name Release Date Release Note Reference
Krutrim-2-Base 2024-01-31 Trained with MN12B architecture
Krutrim-2-Instruct 2024-01-31 Finetuned and aligned version of Krutrim-2-Base Here

Data Freshness

  • The dataset includes information up to April 2024.

Model Architecture

  • Layers: 40
  • Hidden Dimension: 5,120
  • Head Dimension: 128
  • Activation Function: SiLU
  • Number of Heads: 32
  • Number of KV-Heads: 8 (GQA)
  • Rotary Embeddings: Theta = 1M
  • Vocabulary Size: 131072 (2^17)
  • Architecture Type: Transformer Decoder (Auto-regressive Language Model)

Evaluation Results

English/Code/Math Benchmarks

We use the LM Evaluation Harness to evaluate our model on the En benchmarks tasks. Please note that at the time of writing this report, we were unable to use the evaluation framework for llama-3.3-70B, Gemini-1.5-flash and GPT-4o. We currency report the available published numbers for these models. We realise that the prompt templates and few-shot settings might vary and are working to make these evaluations consistent.

Benchmark Krutrim-1-7B MN-12B-Instruct Krutrim-2-12B llama-3.3-70B Gemini-1.5 Flash GPT-4o
Hellaswag (0-shot) - Accuracy 0.74 0.82 0.83 0.95 0.87 (10-shot) 0.95 (10-shot)
Winogrande (0-shot) - Accuracy 0.67 0.74 0.77 0.85 (5-shot) - 0.88 (5-shot)
OpenBookQA (0-shot) - Accuracy 0.45 0.46 0.49 - - -
CommonSenseQA (0-shot) - Accuracy 0.74 0.70 0.74 - - 0.85
TruthfulQA (0-shot) - Accuracy 0.49 0.54 0.59 - - 0.59
MMLU (5-shot) - Accuracy 0.47 0.68 0.63 0.82 0.79 0.86
TriviaQA (5-shot) - EM 0.44 0.72 0.62 - - -
NaturalQuestions (5-shot) - EM 0.15 0.28 0.26 - - -
GSM8K (0-shot) - EM 0.07 0.74 0.71 0.93 (8-shot, CoT) 0.86 (11-shot) 0.89
ARC_Challenge (0-shot) - Accuracy 0.48 0.59 0.60 0.93 (25-shot) - 0.50
ARC_Easy (0-shot) - Accuracy 0.73 0.80 0.82 - - -
HumanEval - Pass@10 0.00 0.23 0.80 0.88 0.74 (0-shot) 0.90
IF_Eval (0-shot) - Accuracy 0.27 0.46 0.73 0.92 - 0.84

Indic Benchmarks

Benchmark Metric Krutrim-1-7B MN-12B-Instruct Krutrim-2-12B llama-3.3-70B Gemini-1.5 Flash GPT-4o
IndicSentiment (0-shot) Accuracy 0.65 0.70 0.95 0.96 0.99 0.98
IndicCOPA (0-shot) Accuracy 0.51 0.58 0.80 0.83 0.88 0.91
IndicXParaphrase (0-shot) Accuracy 0.67 0.74 0.88 0.87 0.89 0.91
IndicXNLI (0-shot) Accuracy 0.47 0.54 0.55 0.61 0.70 0.75
IndicQA (0-shot) Bert Score 0.90 0.90 0.91 0.89 0.94 0.93
CrossSumIN (1-shot) chrF++ 0.04 0.17 0.21 0.26 0.24 0.24
FloresIN Translation xx-en (1-shot) chrF++ 0.54 0.50 0.58 0.60 0.62 0.63
FloresIN Translation en-xx (1-shot) chrF++ 0.41 0.34 0.48 0.46 0.47 0.48
IN22 Translation xx-en (0-shot) chrF++ 0.50 0.48 0.57 0.58 0.55 0.60
IN22 Translation en-xx (0-shot) chrF++ 0.36 0.33 0.45 0.42 0.44 0.44

BharatBench

The existing Indic benchmarks are not natively in Indian languages, rather, they are translations of existing En benchmarks. They do not sufficiently capture the linguistic nuances of Indian languages and aspects of Indian culture. Towards that Krutrim released BharatBench - a natively Indic benchmark that encompasses the linguistic and cultural diversity of the Indic region, ensuring that the evaluations are relevant and representative of real-world use cases in India.

Benchmark Metric Krutrim-1-7B MN-12B-Instruct Krutrim-2-12B llama-3.1-70B Gemma-2-27B GPT-4o
Indian Cultural Context (0-shot) Bert Score 0.86 0.56 0.88 0.88 0.87 0.89
Grammar Correction (5-shot) Bert Score 0.96 0.94 0.98 0.98 0.96 0.97
Multi Turn (0-shot) Bert Score 0.88 0.87 0.91 0.90 0.89 0.92
Multi Turn Comprehension (0-shot) Bert Score 0.90 0.89 0.92 0.93 0.91 0.94
Multi Turn Translation (0-shot) Bert Score 0.85 0.87 0.92 0.91 0.91 0.92
Text Classification (5-shot) Accuracy 0.61 0.71 0.76 0.88 0.86 0.89
Named Entity Recognition (5-shot) Accuracy 0.31 0.51 0.53 0.61 0.65 0.65

Qualitative Results

Below are the results from manual evaluation of prompt-response pairs across languages and task categories. Scores are between 1-5 (higher the better). Model names were anonymised during the evaluation.

cumulative_score_category cumulative_score_langauge

Usage

To use the model, you can load it with AutoModelForCausalLM as follows:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "krutrim-ai-labs/Krutrim-2-instruct"

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Add custom chat template
prompt_dict = [{"role":'system','content':"You are an AI assistant."},{"role":'user','content':"Who are you?"}]
prompt = tokenizer.apply_chat_template(prompt_dict, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(prompt, return_tensors='pt')
inputs.pop("token_type_ids", None)

# Generate response
outputs = model.generate(
    **inputs,
    max_length=4096,
    temperature=0.3
)

response = tokenizer.decode(outputs[0])

Note: The provided chat template, which is the default chat template, helps generate the best response by structuring conversations optimally for the model. We recommend using temperature=0.3 for the best performance

Limitations

The model was trained on a dataset that includes content from the internet, which may contain toxic language, biases, and unsafe content. The model has undergone extensive post-training to handle these and is continuously evolving. In response to some prompts the model might:

  • amplify biases present in the training data
  • generate toxic responses
  • provide inaccurate, incomplete, or redundant answers
  • generate responses in languages inconsistent with the prompt

Ethical Considerations

  • The model may produce biased or offensive outputs based on its training data.
  • Users should apply human oversight when using the model for decision-making in sensitive areas.
  • While safeguards have been implemented, the model may still generate socially undesirable text in certain contexts.

License

This code repository and the model weights are licensed under the Krutrim Community License.

Contact

Contributions are welcome! If you have any improvements or suggestions, feel free to submit a pull request on GitHub.

Citation 

@misc{Krutrim2LLM2025,
  author = {Aditya Kallappa, Guo Xiang, Jay Piplodiya, Manoj Guduru, Neel Rachamalla, Palash Kamble, Souvik Rana, Vivek Dahiya, Yong Tong Chua, Ashish Kulkarni},
  title = {Krutrim-2 LLM},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/ola-krutrim/Krutrim-2-12B}}
}