A machine translation evaluation pipeline prioritizing Meta's NLLB-200 model and the FLORES-200 dataset, featuring Kaggle-optimized data loading and UMAP embedding visualization.
-
Updated
May 9, 2026 - Jupyter Notebook
A machine translation evaluation pipeline prioritizing Meta's NLLB-200 model and the FLORES-200 dataset, featuring Kaggle-optimized data loading and UMAP embedding visualization.
A LoRA adapter that steers Helsinki-NLP/opus-mt-en-es toward literary prose. 589K trainable parameters — 0.75% of the model — trained on opus_books en-es. 19.99 sacreBLEU against 17.09 for the untouched base model. 2.4 MB; loads on top of the base with PEFT.
🥐 French → English neural machine translation with a Transformer built from scratch in PyTorch: multi-head attention, SentencePiece BPE, multi-GPU training with Accelerate, SacreBLEU evaluation.
English-to-Arabic Translator: neural machine translation with a compact, readable Transformer built from scratch in PyTorch. Works for any language pair: BPE tokenizer, training, beam search, BLEU/chrF evaluation.
From-scratch Transformer NMT for English ↔ Amharic (PyTorch). Data pipeline → BPE → training → beam search → FLORES-200 eval → live demo.
Fine-tuned MarianMT Transformer for English-to-Arabic neural machine translation, evaluated with SacreBLEU and demonstrated with Gradio.
NLP coursework: PPO/RLHF demos (CartPole + TRL) and EN–VI machine translation with GPT/Transformer (from scratch) vs pretrained (GPT-2, MarianMT).
Score machine translation against reference translations with BLEU and chrF, and flag weak segments for review.
Companion repo for A 2×2 Controlled Ablation of Data Quality and Capacity in Transformer MT. Contains the QE-filtered fine-tuning pipeline (Section 7), CometKiwi-22 scoring/filtering scripts, and the full paper LaTeX source. Pretraining lives in Machine_translation
Does fine-tuning NLLB-200 on noisy OPUS-100 help or hurt English→Arabic translation? Single run: in-domain gains (+2.4 BLEU, +1.3 chrF++), out-of-domain losses on FLORES-200 (−2.1 BLEU, −2.0 chrF++), tested with paired bootstrap on five metrics. Includes cleaning pipeline, IBM1 baseline, error analysis.
LoRA fine-tuning of NLLB-200 (1.3B) for Chinese-to-English translation on WMT19, reaching 24.9 BLEU.
BA Thesis Project: Mandarin-to-Cantonese subtitle style transfer for Hong Kong cinema, from corpus construction to fine-tuning and evaluation.
To associate your repository with the sacrebleu topic, visit your repo's landing page and select "manage topics."