From scratch, transparently
Implementing the Transformer directly — rather than fine-tuning a pre-trained system — is the paper's point: a transparent framework whose design and preliminary evaluation can be examined end to end.
The Printing Berlin · room 05 of 11
A Transparent Transformer Framework
Oh, and I have a paper published with Springer.
A from-scratch PyTorch encoder–decoder Transformer for English–Hindi neural machine translation — multi-head self-attention, sinusoidal positional encodings, trained on the IITB parallel corpus — published with Springer Nature at ICTCS 2025 in the LNNS series (volume 1877).
The chapter is titled “A Transparent Transformer Framework for English–Hindi Neural Machine Translation: Design and Preliminary Evaluation”, with an extended version available on Research Square.
Mishra, Y., Singh, N., Rai, A. K., Kumar, S., & Kumar, A. (2025). “A Transparent Transformer Framework for English–Hindi Neural Machine Translation: Design and Preliminary Evaluation.” In ICTCS 2025 — Lecture Notes in Networks and Systems (vol. 1877). Springer Nature. doi.org/10.1007/978-3-032-19672-9_41 ↗
462 accesses at the time of verification · extended version on Research Square
The actual problem
Neural machine translation for English–Hindi is usually reached through pre-trained systems, where the translation quality is real but the mechanism is opaque. Building the Transformer from scratch — attention, positioning, tokenisation and all — makes every design decision inspectable and the evaluation honest about its scope.
The research problem is to design a transparent architecture whose preliminary evaluation is reproducible: a published design with stated data, stated training choices and stated results, at a peer-reviewed venue.
The exploded machine
The architecture is the classic encoder–decoder Transformer implemented directly in PyTorch: multi-head self-attention in both stacks and sinusoidal positional encodings, with no framework-supplied shortcuts hiding the mechanism.
Training and evaluation use the IIT Bombay English–Hindi parallel corpus (IITB), with custom BPE tokenisation preparing the text on both sides.
The chapter reports competitive BLEU scores in its preliminary evaluation (per the Springer abstract), and an extended preprint with fuller analysis is available on Research Square.
Engineering decisions
Sinusoidal positional encodings
Fixed sinusoidal encodings keep the positional mechanism deterministic and inspectable, matching the architecture's transparency goal.
IITB corpus with custom BPE
The IIT Bombay parallel corpus (≈1.6M sentence pairs, resume-sourced) with custom BPE tokenisation grounds the evaluation in a public, citable dataset.
Peer-reviewed publication path
ICTCS 2025 / Springer LNNS v1877 puts the design and its preliminary evaluation through peer review; Research Square carries the extended version.
Proof — verified results
- Publication
- ICTCS 2025 · Springer LNNS vol. 1877verified live
- Chapter DOI
- 10.1007/978-3-032-19672-9_41verified live
- Extended version DOI
- 10.21203/rs.3.rs-9327858/v1verified live
- Headline metrics
- BLEU 49.37 · chrF 78.37 · METEOR 85.89resume-sourced
- Corpus
- IITB English–Hindi parallel corpusper abstract
- Evaluation headline
- “competitive BLEU scores” (per the Springer abstract)per abstract
every number here carries its source — my resume's figures and the synthetic-fixture results are marked for what they are