Logo Lanfrica

w-yidne/hasesa

Domain:

natural language processingeducation

Record type:

project
Creator:
w-y
Host:
Hasesa (ሐሰሳ) — an offline, audio-first Amharic tutor for Ethiopian high-school students. Research groundwork: a curriculum-grounded Amharic/English genetics QA benchmark. # Hasesa · ሐሰሳ **An audio-first, offline Amharic tutor for Ethiopian high-school students.** *Hasesa (ሐሰሳ) — Amharic for "the search / inquiry."* A student speaks a question in Amharic into a budget phone — with no internet — and hears an answer grounded in their official Ministry of Education textbook. No data cost, no account, works on the bus. This repository is the research groundwork for that vision. It starts small on purpose: a rigorous study of the single hardest question underneath the whole idea, done on a laptop with no GPU, before any app is built. --- ## The current study (weeks 1–8) > **How well can phone-sized language models tutor one high-stakes Ethiopian curriculum > topic — and does grounding them in the textbook fix what's broken?** - **Scope:** one chapter, in depth — **Grade 11 Biology, Unit 4: Genetics** (DNA & RNA structure, DNA replication, transcription/translation, cell division, inheritance), English-medium. ~60 curriculum questions built from the unit's own review questions, plus items grounded in a **DNA-extraction lab I ran with Grade 11 students** (see below). - **Why one chapter, not a whole subject:** depth over breadth. Exhaustively benchmarking a single high-stakes unit is more defensible than thin coverage of all of Biology, and it lets the failure-mode analysis go deep. The pipeline generalizes to other units unchanged. - **Experiment:** 4–5 small models (Gemma 3 1B, Llama 3.2 1B, Qwen 2.5 1.5B, + a frontier model as a ceiling), each answering in two conditions — *closed-book* (from memory) vs *retrieval-grounded* (given the relevant textbook passage). - **Also measured:** the "tokenizer tax" — how many more tokens Amharic (fidel script) costs than English across common tokenizers, which is why small on-device models struggle with Amharic. - **Output:** a short technical report + this public repo + reproducible scripts. ### The classroom anchor — a real lab, not a hypothetical The benchmark is tied to teaching I actuall …