EnglishToHindi

SLM — English → Hindi Translation

A small encoder-decoder transformer that translates English into Hindi, trained from scratch on a ~2,000-sentence parallel corpus.

Vanilla "Attention Is All You Need" architecture — no pretrained weights, no frameworks beyond PyTorch, everything from the tokenizer to the BLEU score written out in full.

---

Architecture

The encoder reads the English sentence; the decoder generates Hindi, attending to its own output so far and, through cross-attention, to the whole source.

---

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Prepare the data

python main.py prepare

Splits data/translation.csv into train/val/test and trains the shared BPE tokenizer. The split is grouped by English sentence, so no source sentence appears in two splits, and the tokenizer only ever sees the training half.

3. Train

python main.py train

~60 seconds on a GPU, a few minutes on CPU. The best checkpoint by validation loss lands in checkpoints/best_model.pt.

4. Translate

python main.py translate                        # interactive
python main.py translate -s "Are you at home?"  # one sentence
python main.py translate -s "Good morning" --beam 1   # greedy

5. Benchmark

python main.py benchmark                  # writes benchmarks/benchmark_<timestamp>.txt
python main.py benchmark --cpu            # force CPU
python main.py benchmark --cpu --threads 4
python main.py benchmark --beam 1 4 8     # compare decoding strategies
python main.py benchmark --json out.json  # raw metrics too

---

Results

Held-out test split (90 source sentences), scored against every reference translation available for each sentence:

Teacher-forced on the same split: loss 2.55, perplexity 12.84, top-1 token accuracy 51.2%, top-5 74.7%.

These numbers are low, and honestly so. 1,797 training pairs is a tiny corpus for translation — the model reliably picks up sentence shape, question particles and common phrases, then invents vocabulary it never saw enough of:

EN   Are you at home?
OUT  तुम घर हो क्या?
REF  तुम घर पे हो क्या?

EN   Bad news travels quickly.
OUT  बच्चे सेवार को साथ-कार करो।
REF  बुरी खबर तेज़ी से फैलती है।

Validation loss bottoms out around step 1,200 and rises afterwards while training loss keeps falling — classic overfitting on a small corpus. More parallel data is the only real fix; label smoothing, dropout 0.2 and tied embeddings already do what regularisation can.

---

Metrics

Both scores are implemented in slm/metrics.py, not imported:

• BLEU — n-gram precision up to 4-grams with brevity penalty. Orders with no n-grams (sentences shorter than n) are skipped rather than scored zero.
• chrF — character n-gram F-score, weighted toward recall. Better suited to Hindi than BLEU, since it gives partial credit when the stem is right and only the inflection is wrong.

Where an English sentence has several Hindi translations in the corpus, all of them count as references.

---

Layout

data/translation.csv     English–Hindi parallel corpus
main.py                  CLI entry point
slm/config.py            Model and training hyperparameters
slm/data.py              Corpus loading, cleaning, splitting
slm/tokenizer.py         Shared byte-level BPE tokenizer
slm/prepare_data.py      Split + tokenizer pipeline
slm/dataset.py           Tokenized pairs, padding, batching
slm/model.py             Attention, encoder, decoder, decoding
slm/train.py             Training loop
slm/translate.py         Inference API and CLI
slm/metrics.py           BLEU and chrF
slm/benchmark.py         Benchmark suite and report writer