Two open projects examine how continued pretraining, instruction tuning and domain specialization change capability and alignment in Amharic. Review the agriculture and health scopes, budgets, methods and progress below.
Projects
Current research
Research groundwork underwayAmharic phase
Evaluating alignment and reliability in Amharic agricultural AI
This project measures how Amharic language adaptation and agricultural fine-tuning change model capability, refusal, uncertainty and locally relevant advice.
External funding toward this phase
$0
of $27,500 full target
Groundwork already funded and underway. Addis AI has funded the underlying model research and infrastructure; external funding expands this evaluation phase.
Evaluating alignment and safety in Amharic health AI
This project measures how Amharic language adaptation and health fine-tuning change safety behavior, factuality, uncertainty and escalation.
External funding toward this phase
$0
of $33,000 full target
Groundwork already funded and underway. Addis AI has funded the underlying model research and infrastructure; external funding expands this evaluation phase.
How does each adaptation stage change capability and alignment in Amharic?
The same evaluation suite follows the model through four checkpoints. This separates changes from language learning, instruction tuning and domain specialization.
Stage 1
Aligned instruction-tuned open checkpoint
Start from an instruction-tuned open model with an existing safety and instruction-following baseline. The exact checkpoint and configuration will be selected and frozen in the public protocol before the main evaluation begins.
Stage 2
Amharic language adaptation
Apply the frozen Amharic language-adaptation configuration, then measure capability, factuality, language, local relevance, safety and alignment behavior again.
Stage 3
Broad Amharic instruction tuning
Teach general Amharic instruction following, then repeat the same evaluation suite.
Stage 4
Domain specialization
Apply agriculture or health supervised fine-tuning, then compare the change from every earlier stage.
At every stage: capability and safety evaluation, failure analysis, mitigation where indicated, and open results and artifacts.
What Open Research is
Public research scopes with methods, budgets and updates.
Each project has a defined question, evaluation plan, funding target, minimum scope and release plan. The project pages show what exists, what still needs funding and what the work can establish.
These projects use agriculture and health as experimental settings. The broader contribution is evidence about how alignment changes during low-resource language adaptation.
What already exists
Public models, tools and benchmark methods
Addis AI has previously carried out vocabulary extension, continued pretraining and supervised fine-tuning for Amharic.
That work informs the adaptation method, but the agriculture and health studies use a separate experimental configuration that will be frozen in a public protocol.
The research scopes and funding records live on addisai.ch. Public models, datasets, tools and benchmarks remain on Addis AI's open-source sites.
Open models, public benchmark results and research manuscripts from Addis AI. Release and manuscript status are reported separately.
Available in the API
Funded by Addis AI
Addis Scribe for Afaan Oromo
Afaan Oromo speech-to-text for uploads and live audio
Addis Scribe now transcribes Afaan Oromo (language code om) alongside Amharic through the same API, with file uploads, live streaming, word timestamps, SRT and VTT captions, and speaker labels. The model was trained by Addis AI.
Public benchmark result: On 150 held-out WAXAL Afaan Oromo test clips of natural speech (23 speakers, 45 minutes, human transcripts), Addis Scribe records 21.1% WER and 6.8% CER live, the lowest error rate of every system we tested and half the errors of Google's live Oromo recognition (42.4% WER). The best open model tested, Ethio-ASR Oromo, records 24.8% offline. Final text arrives about 40 ms after the speaker stops, compared with about 1 s for Google. On the full WAXAL Oromo test split (3,782 clips), Addis Scribe uploads record 20.8% WER. The references, every system's transcripts, the scorer and the scripts are public, so you can rerun each system with your own API keys and check these numbers.
Scope note: WAXAL is one public corpus of natural speech. Latency was measured with audio fed at real-time speed; Google's figures include the network round trip.
An open Amharic automatic speech recognition checkpoint adapted from Qwen3-ASR-0.6B with a broader speech-text mixture for stronger generalization beyond the curated training domain.
Public benchmark result: On the public fixed held-out evaluations, the broader mixed checkpoint improved WAXAL word error rate by 0.99 percentage points over the curated variant. The curated checkpoint remained 0.47 points better on its source-matched heldout.
Scope note: The reported benchmarks cover limited held-out corpora and do not represent universal Amharic speech recognition performance.
Vocabulary extension, continued pretraining and supervised fine-tuning
A technical study of adapting a small open foundation model to Amharic through vocabulary extension, continued pretraining and supervised fine-tuning. The work documents the adaptation pipeline, data-engineering lessons, negative results and evaluation-methodology questions that motivate the next controlled experiments.
Scope note: This prior work informs the adaptation method. The controlled agriculture and health studies described here are separate experiments.
Evaluating a Production Hausa Speech-to-Text System
Accuracy, robustness and latency under domain-specific speech conditions
A research manuscript evaluating a deployed Hausa speech-recognition system with a reproducible benchmark focused on accuracy, robustness and operational performance under clean and noisy conditions.
Scope note: This evaluation manuscript belongs to the broader speech-research portfolio and is not evidence for the agriculture or health alignment studies.
Automated measures cannot decide whether Amharic is natural, advice fits local conditions, a red flag was missed or uncertainty was stated well. Native speakers and domain reviewers define the scenario set, score outputs and resolve disagreements.
What funding enables
Native Amharic evaluators
Agronomists and qualified health professionals
Scenario creation, contextual tests and preference labels
Red-team cases, repeated runs and mitigation tests
Reproducibility work and public release
Open-source commitment
Methods and results will be public.
We will publish protocols, taxonomies, scoring code, rubrics, stage comparisons, aggregate and per-slice results, mitigation results, configurations, limitations and reproduction instructions. We will publish null and negative results.
Datasets, checkpoints and source material will be released where privacy, licensing and safety permit.
Research updates
We publish progress, changes in scope, negative results and delays as they happen.
2026-08-31
Agriculture: Project scope published
The public scope, stage comparison, initial budget and planned outputs are available for review.
2026-08-31
Health: Project scope published
The public scope, stage comparison, initial budget and planned outputs are available for review.
Funding transparency
Public totals come from the project ledger.
Funding totals include only cleared contributions and confirmed project-specific grants or sponsorships.
The funding policy covers project use, minimum scope, scope changes, overfunding, unused funds, fees, reconciliation and contributor terms.
