Phi-4-Hindi

Phi-4-Hindi is a 14.7B parameter pre-trained and instruction-tuned bilingual large language model for both Hindi and English, trained on a mixed language dataset.

• ~1% better performance on English Tasks compared to the original (average benchmark scores)
• ~4% better performance on Hindi Tasks compared to the original (average benchmark scores)
• ~10% less emissions than the original (as reported on benchmark evaluations like open-llm-leaderboard)
• Less Biases due to ordering of choices while answering MCQs

Model Details:

• Developed by: Traversaal.ai, 1-800-LLMs
• Language(s) (NLP): Optimized for Hindi and English
• License: Apache 2.0
• Paper : TBA April 15

Intended Use

We release Phi-4-Hindi under the Apache 2.0 license, encouraging researchers, developers, and enterprises to experiment with and build upon the model, particularly for bilingual, multilingual and non-English applications. At the time of release, the model demonstrated state-of-the-art performance across an extensive English and Hindi evaluation suite.

Some potential downstream applications are as follows:

• Research: This model serves as a valuable tool for researchers and developers working in NLP.
• Commercial Use: It can be utilized as a foundational model for fine-tuning to meet specific industry needs.
  Possible applications include:
  • AI-powered Chat Assistants
  • Customer Support Service
  • Educational tools for language learning

Target audiences who may benefit from our model:

• Academics: Researchers focused on Hindi and multilingual NLP advancements.
• Businesses: Companies catering to Hindi-speaking and bilingual users.
• Developers: Those integrating Hindi language capabilities into applications and services.
• Educational Institutions: Schools and universities developing AI-powered learning tools.

Prompt Formats

Task	Input Format
Natural Language Inference	"Text1 ### Text2 ### NLI ###"
Multiple Choice Questions	"Question ### A) a, B) b,... ### MCQ ###"
Numeric Questions	"Question ### NUMERIC ###"
Boolean Questions	"Question ### BOOLEAN ###"
Questions seeking Long responses	"Question ### LONG RESPONSE ###"
Short responses (few words)	"Input ### DIRECT RESPONSE ###"
Coding	"Input ### CODE ###"
Text Summarization	"Input ### SUMMARIZE ###"
Paraphrasing/Rephrasing	"Input ### PARAPHRASE ###"
Translation to specified language	"Input ### TRANSLATION [lang] ###"
Text Simplification/ELI5	"Input ### SIMPLIFY ###"

The following prompt formats were used during training and are better suited for usage, however the model works well even without such formatting

Out-of-Scope Use

While Phi-4-Hindi is a powerful bilingual model designed for Hindi and English, it is crucial to acknowledge its limitations and the potential for misuse. The model must not be used in ways that violate any applicable laws or regulations. Below are specific scenarios where its use is restricted:

• Harmful or Malicious Use: The model should not be employed to create or distribute harmful, misleading, or inappropriate content, including but not limited to:

  • Encouraging hate speech, violence, or discrimination
  • Spreading misinformation or false narratives
  • Facilitating or promoting illegal activities

• Sensitive Data Handling: The model is not designed to process or generate personal, confidential, or sensitive information.

• Language Constraints: While optimized for Hindi and English, the model should not be assumed to have the same proficiency in other languages.

• High-Risk Decision-Making: It should not be used for critical decision-making without human oversight, especially in medical, legal, financial, or safety-related contexts.

Bias, Risks, and Limitations

While efforts have been made to minimize biases, it is likely that the model, as with all LLM models, will exhibit some bias.

The model is trained as an AI assistant for Hindi and English speakers. The model is limited to produce responses for queries in these two languages and may not produce appropriate responses to other language queries.

By using this model, you acknowledge and accept that, as with any large language model, it may generate incorrect, misleading and/or offensive information or content. The information is not intended as advice and should not be relied upon in any way, nor are we responsible for any of the content or consequences resulting from its use. We are continuously working to develop models with greater capabilities, and as such, welcome any feedback on the model~~

Evaluation:

We evaluated our models on multiple well-known benchmarks to measure their effectiveness against other leading models, and the results are as follows:

Model	ARC-C	ARC-E	BoolQ	CMCQ	MMLU	Average*	MMLU-Pro	GPQA	MuSR	BBH	MATH-Hard
AryaBhatta-GemmaUltra-8.5B	22.70	25.04	22.95	62.23	23.70	31.32	22.66	25.34	42.72	41.12	2.95
Airavata-7B	25.09	30.47	25.31	62.17	33.20	35.25	16.35	27.43	37.57	36.00	13.60
sarvam-1-2B	30.03	33.25	62.17	42.80	27.90	39.23	-	-	-	-	-
Nemotron-4-Mini-Hindi-Instruct	55.80	71.63	62.11	68.10	43.20	60.17	25.95	30.87	41.53	40.11	2.04
Llama-3-Nanda-10B-Chat	65.36	80.64	82.29	67.60	50.61	69.30	31.57	30.12	43.52	49.38	5.59
Krutrim-2-12b-instruct	67.32	81.10	84.74	76.30	56.10	73.11	-	-	-	-	-
aya-expanse-8b	74.06	87.08	86.45	83.30	56.89	77.56	30.04	30.29	37.17	49.42	7.02
aya-expanse-32B	85.41	95.08	90.43	89.80	69.71	86.08	41.30	32.55	38.62	56.29	13.37
Our Qwen Model (14b)	90.61	94.82	88.53	90.70	75.00	87.93	52.63	36.24	44.84	64.97	25.08
Our Phi Model (14b)	97.39	92.24	87.65	87.40	75.59	88.05	52.39	39.77	49.07	66.97	23.11

Table 1: Metrics (.2f) of our models and other LLMs over several English benchmarks

Model	ARC-C	ARC-E	BoolQ	CMCQ	MMLU	Average
AryaBhatta-GemmaUltra-8.5B	22.70	25.08	22.95	62.17	23.80	31.34
Airavata-7B	22.87	25.13	23.28	62.17	33.20	33.33
sarvam-1-2B	32.76	35.06	62.16	47.10	24.22	40.26
Llama-3-Nanda-10B-Chat	45.99	60.56	71.96	54.70	36.35	53.91
Nemotron-4-Mini-Hindi-4B-Instruct	50.68	63.72	68.74	51.30	37.18	54.32
Krutrim-2-12b-instruct	56.83	70.66	78.86	64.10	46.51	63.39
aya-expanse-8b	57.42	72.90	80.42	69.00	43.39	64.63
aya-expanse-32B	73.29	85.48	87.73	79.70	56.96	76.63
Our Qwen Model (14b)	74.06	81.23	84.07	78.20	53.85	74.82
Our Phi Model (14b)	81.74	89.06	86.02	78.70	56.39	78.38

Table 2: Metrics (.2f) of our models and other LLMs over several Hindi benchmarks

Benchmark	Lang	Qwen-2.5-14B-Instruct	Our Qwen	Change	Phi-4	Our Phi	Change
ARC-Easy	En	95.45	94.82	🔻 0.63	97.31	97.39	🔼 0.08
	Hi	78.49	81.23	🔼 2.74	86.87	89.06	🔼 2.19
ARC-Challenge	En	90.87	90.61	🔻 0.26	92.41	92.24	🔻 0.17
	Hi	69.62	74.06	🔼 4.44	79.18	81.74	🔼 2.56
BoolQ	En	86.09	88.53	🔼 2.44	86.30	87.65	🔼 1.35
	Hi	78.89	84.07	🔼 5.18	82.72	86.02	🔼 3.30
Context-MCQ	En	91.20	90.70	🔻 0.50	86.30	87.40	🔼 1.10
	Hi	77.40	78.20	🔼 0.80	75.70	78.70	🔼 3.00
MMLU	En	74.37	75.00	🔼 0.63	74.67	75.59	🔼 0.92
	Hi	52.16	53.85	🔼 1.69	53.24	56.39	🔼 3.15
Average	En	87.60	87.93	🔼 0.33	87.40	88.05	🔼 0.65
	Hi	71.31	74.82	🔼 3.51	75.54	78.38	🔼 2.84
Overall		79.46	81.38	🔼 1.92	81.47	83.22	🔼 1.75

Table 3: Performance of our models compared to originals over each benchmark : evals through log likelihoods

Benchmark	Lang	Qwen-2.5-14B-Instruct	Our Qwen	Change	Phi-4	Our Phi	Change
MMLU-Pro	En	49.04	52.63	🔼 3.59	53.78	52.39	🔻 1.39
MATH hard	En	00.00	25.08	N/A	12.31	23.11	🔼 10.80
GPQA	En	32.21	36.24	🔼 4.03	33.72	39.77	🔼 6.05
MuSR	En	40.87	44.84	🔼 3.97	41.01	49.07	🔼 8.06
BigBench-Hard	En	63.74	64.97	🔼 1.23	68.60	66.97	🔻 1.63
Average		37.17	44.75	🔼 7.58	41.88	46.26	🔼 4.38

Table 4: Performance of our models compared to originals over each benchmark : evals through eval-harness

Recommendations

It is advisable for users to:

• Refrain from deploying the model in sensitive domains without human supervision.
• Cross-check factual information generated by the model for accuracy.
• Continuously assess the model to ensure compliance with ethical standards.
• Be mindful of potential biases and unintended outputs, especially in critical applications.

Emissions

We belive our usage of shorter and compressed instruction-reponse pairs in training resulted in the model responding in simplified manner while meeting the requirements/ arriving at the correct answers. Hence the better scores while reducing emissions.

Unlike distillation from reasoining or CoT models which produced unnecessarily long responses like "Next we proceed with...", "Ok lets do this...." during generation of step by step solutions of a math problem, we use only the step by step math part ignoring such fillers, for datasets with multiple step-by-step solutions which are correct, we chose the shortest one to train our models.

Model Responses vs Order of Choices in MCQs

As benchmarks like MMLU-Pro have upto 10 choices, while most training datasets consist of typically 4-5 choices, we modified the ordering and labelling of choices i.e re-ordering choices to create an imbalance opposing the original model's choice distribution, replacement of labels from A/B/C/D to a/b/c/d or 1/2/3/4 or w/x/y/z etc.. in 5% of the MCQ samples for better robustness This resulted in less bias towards the earlier choices among MCQs as compared to the original phi-4. The below images are a distution of choices selected by the model while being evaluated over MMLU-pro

Team

• Ram Mohan Rao Kadiyala
• Siddartha Pullakhandam
• Siddhant Gupta
• Drishti Sharma
• Jebish Purbey
• Kanwal Mehreen
• Muhammad Arham
• Hamza Farooq

Correspondence