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.
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.
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:
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2import torch
34model_id ="krutrim-ai-labs/Krutrim-2-instruct"56# Load model and tokenizer7model = AutoModelForCausalLM.from_pretrained(model_id)8tokenizer = AutoTokenizer.from_pretrained(model_id)910# Add custom chat template11prompt_dict =[{"role":'system','content':"You are an AI assistant."},{"role":'user','content':"Who are you?"}]12prompt = tokenizer.apply_chat_template(prompt_dict, add_generation_prompt=True, tokenize=False)13inputs = tokenizer(prompt, return_tensors='pt')14inputs.pop("token_type_ids",None)1516# Generate response17outputs = model.generate(18**inputs,19 max_length=4096,20 temperature=0.321)2223response = 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.
