Research & Projects

Research benchmarks and empirical investigations across low-resource Semitic NLP, Ethiopic LLM tokenization, retrieval architectures, parameter-efficient fine-tuning (PEFT), and dialect speech recognition.

Large Language Models & Semitic NLP

Amharic Tokenization Benchmark

LLM Tokenization FLORES-200 Token Tax

Quantifying the 9.6x cross-lingual Token Tax and byte-fallback collapse across 9 LLMs on FLORES-200.

Details & Findings

Amharic LLM RAG Benchmark

RAG Retrieval Hybrid RRF Dagim QA

Empirical evaluation of Dense, BM25, and Hybrid RRF retrieval for Amharic Question Answering on Dagim QA.

Details & Findings

Parameter-Efficient Amharic LoRA

LoRA / PEFT Qwen 2.5 Bits-per-Character

LoRA adaptation of Qwen2.5-0.5B on Amharic Wikipedia; evaluated with Bits-per-Character and paired bootstrap CIs.

Details & Findings

Subword Tokenization & Fidel Normalization

WordPiece vs BPE Fidel Normalization MasakhaNEWS

WordPiece vs. Byte-Level BPE vs. Unigram on MasakhaNEWS with orthographic Fidel homophone merging.

Details & Findings

Dialect Speech Recognition & Audio AI

Whisper ASR for Wollo Amharic Dialect

OpenAI Whisper PEFT / LoRA Wollo Dialect Dialect ASR

Parameter-efficient fine-tuning (PEFT) of OpenAI Whisper for regional Amharic dialect speech recognition.

Details & Findings