Indic-gemma-2b-finetuned-sft-Navarasa

Indic-gemma-2b-finetuned-sft-Navarasa

This model is based on google/gemma-2b and hase been LoRA finetuned on 9 Indian languages and English language instruction datasets:

1. Hindi - ravithejads/samvaad-hi-filtered, HydraIndicLM/hindi_alpaca_dolly_67k(sampled)

2. Telugu - Telugu-LLM-Labs/yahma_alpaca_cleaned_telugu_filtered_and_romanized, Telugu-LLM-Labs/teknium_GPTeacher_general_instruct_telugu_filtered_and_romanized

3. Tamil - abhinand/tamil-alpaca

4. Kannada - Tensoic/airoboros-3.2_kn, Tensoic/gpt-teacher_kn

5. Malayalam - VishnuPJ/Alpaca_Instruct_Malayalam

6. Gujarati - Tensoic/Alpaca-Gujarati

7. Punjabi - HydraIndicLM/punjabi_alpaca_52K

8. Bengali - HydraIndicLM/bengali_alpaca_dolly_67k(alpaca filtered)

9. Odia - OdiaGenAI/Odia_Alpaca_instructions_52k, OdiaGenAI/gpt-teacher-roleplay-odia-3k

10. English - yahma/alpaca-cleaned

The model is finetuned using unsloth library and we provide inference code using the same for faster inference. Alternatively you can use HuggingFace Library for inference.

Training Details:

The model is trained on approx 500K instruction samples.

1. GPU: 1 A100, 80GB
2. Time: 36.5 Hours
3. Platform: E2E Networks

Installation

!pip install "unsloth[colab-ampere] @git+https://github.com/unslothai/unsloth.git"

Input Text Format

### Instruction: {instruction}

### Input: {input}

## Response: {response}

Inference With Unsloth

from unsloth import FastLanguageModel
import torch
max_seq_length = 2048
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = False 
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Telugu-LLM-Labs/Indic-gemma-2b-finetuned-sft-Navarasa",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    device_map="auto"
)
FastLanguageModel.for_inference(model) # Enable native 2x faster inference

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

### Input:
{}

### Response:
{}"""

input_text = input_prompt.format(
        "Tranlsate following sentence to Hindi.", # instruction
        "This model is developed by Telugu LLM Labs", # 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)

Inference with HuggingFace

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained(
    "Telugu-LLM-Labs/Indic-gemma-2b-finetuned-sft-Navarasa",
    load_in_4bit = False,
    token = hf_token
)
tokenizer = AutoTokenizer.from_pretrained("Telugu-LLM-Labs/Indic-gemma-2b-finetuned-sft-Navarasa")

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

### Input:
{}

### Response:
{}"""

input_text = input_prompt.format(
        "Tranlsate following sentence to Hindi.", # instruction
        "This model is developed by Telugu LLM Labs", # 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]

Refer to the blog post for sample examples.

Please check our Code Repositoryfor training and inference scripts.

Developers:

The model is a collaborative effort by Ravi Theja and Ramsri Goutham. Feel free to DM either of us if you have any questions.