AryaBhatta-GemmaOrca-Merged

This model is a part of two model series, AryaBhatta-1 and AryaBhatta-2 and is finetuned from HuggingFaceH4/zephyr-7b-gemma-v0.1 or Google/gemma and is finetuned on 9 Indian languages (Hindi, Tamil, Punjabi, Bengali, Gujarati, Oriya, Telugu, Kannada, Malayalam) plus English.

There are two models. One finetuned on Google's Gemma and one fine-tuned on Zephyr's Gemma base. Repo for other one (Zephyr one): GenVRadmin/AryaBhatta-GemmaOrca-2-Merged

To improve the resoning and maths skills, we first SFT tune the gemma on Microsoft's Orca datasets.

We utilize Orca maths Hindi dataset: GenVRadmin/Aryabhatta-Orca-Maths-Hindi
And original Orca maths dataset: microsoft/orca-math-word-problems-200k

This pushes the MATHS score from 24.3 in Gemma-7B to 25.5 in Zephyr-Gemma and 31.6 in GemmaOrca.

The model is then finetuned on GenVR's Samvaad datasets (GenVRadmin/Samvaad-Indic-Positive and GenVRadmin/Samvaad-Tamil-Mixtral and a subset of GenVRadmin/Samvaad-Mixed-Language-3).

This is then finetuned on various open sourced datasets like:

Telugu-LLM-Labs/yahma_alpaca_cleaned_telugu_filtered_and_romanized
Telugu-LLM-Labs/teknium_GPTeacher_general_instruct_telugu_filtered_and_romanized
abhinand/tamil-alpaca
Tensoic/airoboros-3.2_kn
Tensoic/gpt-teacher_kn
Tensoic/Alpaca-Gujarati
HydraIndicLM/bengali_alpaca_dolly_67k
Open-Orca/OpenOrca
pankajmathur/alpaca_orca
OdiaGenAI/Odia_Alpaca_instructions_52k
OdiaGenAI/gpt-teacher-roleplay-odia-3k
GenVRadmin/Samvaad-Punjabi-Mini
pankajmathur/WizardLM_Orca

The model achieves following scores on benchmarks:

Model AGIEval GPT4All TruthfulQA BigBench Average ⬇️
AryaBhatta-GemmaOrca 35.9 72.26 53.85 40.35 50.59
zephyr-7b-beta 37.52 71.77 55.26 39.77 51.08
zephyr-7b-gemma-v0.1 34.22 66.37 52.19 37.10 47.47
mlabonne/Gemmalpaca-7B 21.6 40.87 44.85 30.49 34.45
google/gemma-7b-it 21.33 40.84 41.70 30.25 33.53

How to use:-

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained(
    "GenVRadmin/AryaBhatta-GemmaOrca",
    load_in_4bit = False,
    token = hf_token
)
tokenizer = AutoTokenizer.from_pretrained("GenVRadmin/AryaBhatta-GemmaOrca")

input_prompt = """
### Instruction:
{}

### Input:
{}

### Response:
{}"""

input_text = input_prompt.format(
        "Answer this question about India.", # instruction
        "Who is the Prime Minister of India", # input
        "", # output - leave this blank for generation!
    )

inputs = tokenizer([input_text], return_tensors = "pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens = 300, use_cache = True)
response = tokenizer.batch_decode(outputs)[0]