Description

Today we learn how the LLM translate to any language for free with local Large language models.

1. Model: Helsinki-NLP/opus-mt-en-hi

This is a MarianMT (Marian Machine Translation) model.

It was trained on the OPUS dataset (a large collection of parallel corpora).

This particular checkpoint is English → Hindi.

The architecture is based on Transformer Seq2Seq (encoder–decoder).

2. Code Walkthrough

a. Load Tokenizer

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

Tokenizer converts human text (“I am going to school.”) into tokens (numbers) that the model can understand.

Example: “I am going to school.” → [1234, 567, 890, …]

It also handles things like lowercasing, splitting into subwords (Byte Pair Encoding).

b. Load Model

model = AutoModelForSeq2SeqLM.from_pretrained(model_name, trust_remote_code=True)

Loads the encoder-decoder translation model.

Encoder reads English tokens → creates context embeddings.

Decoder takes those embeddings and generates Hindi tokens step by step.

c. Prepare Input

inputs = tokenizer(text, return_tensors=”pt”, padding=True)

Converts text into PyTorch tensors (input_ids, attention_mask).

Example:

input_ids: [37, 14, 567, 2021, …]

attention_mask: [1, 1, 1, 1, …] (marks real tokens vs padding).

d. Generate Translation

outputs = model.generate(**inputs, max_new_tokens=50)

The generate method runs beam search / greedy decoding to predict Hindi tokens one by one.

Example:

• Step 1: → predicts "मैं"

• Step 2: “मैं” → predicts “स्कूल”

• Step 3: “स्कूल” → predicts “जा”

• Step 4: “जा” → predicts “रहा हूँ।”

Continues until </s> (end-of-sequence token) is generated.

e. Decode Back to Text

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Converts model’s token IDs back into human-readable Hindi text.

Output for “I am going to school.” would be something like:

मैं स्कूल जा रहा हूँ।

3. Theory Recap

Input text → Tokenization → Encoder → Context embeddings

Decoder → Predicts Hindi tokens step by step using context + attention

Beam search / greedy decoding → Generates most likely translation

Tokenizer.decode → Converts tokens back into Hindi text

Here is code to use LLM for language translation

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "Helsinki-NLP/opus-mt-en-hi"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, trust_remote_code=True)

text = "I am going to school."
print("Input")
print(text)
inputs = tokenizer(text, return_tensors="pt", padding=True)
outputs = model.generate(**inputs, max_new_tokens=50)
print("translated text")
print(tokenizer.decode(outputs[0], skip_special_tokens=True))