Sankhaya_Indic_ITN

🔧 Inference Example

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
import re

MODEL_PATH = "panchajanya-ai/Sankhaya_Indic_ITN"
FINETUNED_ADAPTER_PATH = "panchajanya-ai/Sankhaya_Indic_ITN"

# Load model & tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.float32,
    device_map="cpu"
)

try:
    model = PeftModel.from_pretrained(base_model, FINETUNED_ADAPTER_PATH)
except Exception:
    model = base_model

model.eval()

alpaca_prompt = """### Instruction:
Convert the Hindi number words to numeric form.

### Input:
{}

### Response:
"""

# Define common Hindi number word to digit map
with open("hindi_number_mappings.txt", "r", encoding="utf-8") as file:
    file_content = file.read()

# Safely evaluate the string content as a Python dictionary
STATIC_NUM_WORD_MAP = ast.literal_eval(file_content)

# Extract all the keys (Hindi number words) into HINDI_NUM_WORDS
HINDI_NUM_WORDS = list(STATIC_NUM_WORD_MAP.keys())

# Extract number word chunks from sentence
def extract_all_number_chunks(sentence):
    words = sentence.split()
    chunks, temp = [], []
    for word in words:
        if word in HINDI_NUM_WORDS or re.match(r"^[०-९]+$", word):
            temp.append(word)
        else:
            if temp:
                chunks.append(" ".join(temp))
                temp = []
    if temp:
        chunks.append(" ".join(temp))
    return chunks

# Run LLM if not static mapped
def convert_digit_words_to_number(chunk):
    # First check static map
    if chunk in STATIC_NUM_WORD_MAP:
        return STATIC_NUM_WORD_MAP[chunk]

    # Else ask model
    prompt = alpaca_prompt.format(chunk)
    inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
    outputs = model.generate(
        **inputs,
        max_new_tokens=10,
        use_cache=True,
        eos_token_id=tokenizer.eos_token_id
    )
    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
    match = re.search(r"Response:\s*(\d+)", decoded)
    return match.group(1) if match else None

# Main replacement function
def remove_duplicates(s):
            seen = set()
            result = []
            for char in s:
                if char not in seen:
                    seen.add(char)
                    result.append(char)
            return ''.join(result)
def ITN(sentence: str) -> str:
    chunks = extract_all_number_chunks(sentence)
    for chunk in chunks:
        numeric = convert_digit_words_to_number(chunk)
        if numeric:
            sentence = sentence.replace(chunk, numeric)
    return sentence

result = ITN("कृष्ण को पैंतालीस हजार बानवे रुपये भेजें")
print(result)