language:
- am
- en
license: cc-by-4.0
task_categories:
- text-generation
- question-answering
tags:
- amharic
- ethiopian
- instruction-tuning
- nlp
- african-languages
- low-resource
- culture
- orthodox-christianity
- diaspora
- conversational
pretty_name: AddisGPT Amharic Instruction Dataset
size_categories:
- n<1K
AddisGPT-Amharic-Instruction
A human-verified, fully conversational Amharic instruction-tuning dataset sourced entirely from real AddisGPT user interactions.
796 curated instruction–output pairs spanning 14 topics, drawn exclusively from anonymized conversations with AddisGPT — an Amharic-first AI assistant serving Ethiopian and diaspora communities. Every pair is an organic user question paired with the assistant's response; there is no synthetic, templated, or third-party content.
Dataset Summary
| Attribute | Value |
|---|---|
| Total pairs | 796 |
| Format | Conversational (organic user prompt → assistant response) |
| Languages | Amharic (681 · 85.6%), English (115 · 14.4%) |
| Topics | 14 |
| Avg. output length | 972 characters |
| Quality scores | 4–5 (human-verified; see scoring section) |
| Cultural relevance | 89.7% Ethiopian-contextualized (cultural_score ≥ 1) |
| Source | Real AddisGPT user conversations (anonymized) — 100% first-party |
| Date range | 2026-03-25 to 2026-08-28 |
| Creator | TBS Technologies (Virginia, USA) |
Quickstart
from datasets import load_dataset
dataset = load_dataset("tbstechnologies/AddisGPT-Amharic-Instruction")
print(dataset["train"][0])
What this dataset is
Every pair originates from a real person talking to AddisGPT — 451 pairs from anonymous sessions and 345 from signed-in users. These are naturalistic queries: civic and cultural questions, religious guidance, health, personal finance, online work, diaspora and immigration needs, and everyday knowledge, expressed in authentic Amharic (with a minority in English where users chose it). Responses reflect AddisGPT's answers at the time of the conversation.
Because the pairs are conversational rather than templated, they are well suited to general instruction tuning and assistant-style fine-tuning, and they carry genuine Ethiopian cultural grounding that generic multilingual corpora lack.
Topics
| Topic | Count | Description |
|---|---|---|
| General | 368 | Miscellaneous knowledge, mixed everyday questions, greetings, meta questions about the assistant |
| Religion | 74 | Ethiopian Orthodox Christianity, Islam, fasting, holidays, saints, scripture |
| Culture | 64 | Music, coffee ceremony, naming traditions, Amharic poetry, film, customs |
| Politics | 53 | Elections, constitution, governance, ethnic federalism, regional affairs |
| Business | 51 | Online work (Upwork, Fiverr), freelancing, Amazon KDP, business registration |
| History | 38 | Ethiopian emperors, Aksumite civilization, Adwa, modern political history |
| Education | 33 | Ethiopian education system, exams, Amharic language learning, linguistics |
| Technology | 29 | AI, coding, digital literacy, apps, the assistant itself |
| Finance | 22 | Banking, exchange, PayPal availability, tax, rental contracts |
| Health | 20 | Amharic health guidance, traditional medicine, mental health |
| Food | 13 | Ethiopian traditional food, fasting food, nutrition |
| Sports | 11 | World Cup, Premier League, Ethiopian football |
| Geography | 11 | Ethiopian cities, regions, Addis Ababa, neighboring countries |
| Immigration | 9 | Visa processes, US Embassy, passport, diaspora guidance |
Quality Scoring
quality_score is effectively binary in this dataset — pairs are either verified (4) or excellent (5). All rows were human-reviewed and approved.
| Score | Count | Meaning |
|---|---|---|
| 5 | 12 | Excellent — high cultural depth, strong output, exemplary pair |
| 4 | 784 | Verified good — human-reviewed and approved for training |
Cultural Relevance Scoring
Each pair includes a cultural_score (0–5) indicating depth of Ethiopian cultural specificity:
| Score | Meaning | Count |
|---|---|---|
| 5 | Core Ethiopian identity (Aksumite history, Ge'ez alphabet, Fayda ID, coffee ceremony) | 44 |
| 4 | Strongly Ethiopian-contextualized (Ethiopian politics, geography, institutions) | 77 |
| 3 | Amharic-language content with Ethiopian audience value | 88 |
| 2 | Ethiopian-relevant content | 42 |
| 1 | General content in Amharic or with light Ethiopian context | 463 |
| 0 | Non-Ethiopian content (world knowledge asked in Amharic/English) | 82 |
Treat cultural_score as a lower-bound estimate. Scores were assigned via keyword matching, which systematically undercounts semantic cultural depth in Amharic — many pairs carry culturally embedded meaning (proverbs, honorifics, religious or historical allusions) without any explicit trigger word, so their true cultural relevance is higher than the score suggests. Use it as a coarse floor for filtering, not a precise measure.
Fields
| Column | Type | Description |
|---|---|---|
id |
UUID | Unique identifier |
instruction |
text | User prompt / question |
output |
text | Assistant response |
language |
string | amharic or english |
topic |
string | One of 14 topic categories |
quality_score |
int | 4 or 5 (human-verified) |
cultural_score |
int | 0–5 (Ethiopian cultural relevance depth) |
is_verified |
boolean | Human-reviewed and approved (all rows = true) |
source |
string | Origin of the pair (addisgpt_user or anonymous_user) |
dialect |
string | Amharic dialect variant (all standard) |
formality |
string | Register level (all formal) |
turn_count |
int | Conversation turn (all 1 — single-turn pairs) |
created_at |
timestamp | Original conversation date |
Note on constant columns. In this release
dialect(standard),formality(formal),is_verified(true), andturn_count(1) hold a single value across all 796 rows. They are retained deliberately — as documentation of the register/dialect this snapshot represents and as stable placeholders for planned multi-dialect and multi-turn releases where they will carry variance. If your loader or pipeline prefers no zero-variance columns, they are safe to drop (e.g.dataset["train"].remove_columns(["dialect", "formality", "is_verified", "turn_count"])).
Curation Process
This release was produced from AddisGPT's logged user conversations through the following steps:
- Human review and scoring. Each candidate pair was read and assigned a quality score; only pairs scoring 4 or 5 were retained. Low-quality pairs — garbled or single-word prompts, thin decline responses, stale real-time data (exchange rates, weather, live scores), and privacy-risk content (phone numbers, personal identifiers) — were removed.
- First-party only. Earlier working pools mixed in third-party Amharic QA datasets. This release deliberately excludes all external data and contains only pairs originating from real AddisGPT users (
addisgpt_user,anonymous_user), so the entire dataset is TBS Technologies' own to release. - Deduplication. Exact duplicate rows introduced during data export were removed; every
idand every instruction+output pair is unique. - Anonymization. No usernames, account identifiers, or personal data are included.
sourcedistinguishes only signed-in vs. anonymous sessions.
Intended Use
- Fine-tuning LLMs for Amharic instruction-following
- Building Ethiopian-contextualized conversational AI assistants
- Amharic NLP research and benchmarking
- Cross-lingual transfer learning for low-resource African languages
- Ethiopian cultural knowledge grounding
Limitations
- Small size. At 796 pairs this is a focused, high-signal set rather than a large corpus; it is best used for targeted fine-tuning or as a cultural-grounding supplement, not as a sole training source.
- Snapshot answers. Outputs reflect AddisGPT's responses at the time of each conversation and may contain outdated information (political developments, sports results, statistics).
- Topic skew. Coverage follows what Ethiopian users most frequently ask — general, religion, culture, and politics dominate; STEM, legal, and specialized technical topics are underrepresented.
- English minority. 14.4% of pairs are in English, included where users chose English (often for embassy, business, or technical questions).
- Single-turn. All pairs are single-turn; multi-turn dialogue is not represented.
- Cultural scores are a heuristic lower bound. Assigned by keyword matching, not expert annotation; they undercount culturally embedded content that lacks explicit trigger words, so true cultural relevance skews higher than the scores indicate.
Ethical Considerations
- All pairs are derived from anonymized user interactions — no personally identifiable information is included.
- Politically sensitive content (ethnic federalism, regional governance, criticism) is included to reflect authentic Ethiopian discourse, not to endorse any political position.
- Health-related pairs may reference Ethiopian mental health resources (e.g., the 8044 hotline) where appropriate; they are not a substitute for professional medical advice.
- Religious content spans Orthodox Christianity and Islam to reflect Ethiopia's religious diversity.
Citation
@dataset{addisgpt_amharic_instruction_2026,
title={AddisGPT-Amharic-Instruction: A Conversational Amharic Instruction-Tuning Dataset},
author={TBS Technologies},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/datasets/AddisGPT/AddisGPT-Amharic-Instruction}
}
License
This dataset is released under CC BY 4.0. All content originates from TBS Technologies' own AddisGPT product; you are free to share and adapt the material with appropriate attribution.
Contact
- Organization: TBS Technologies
- Product: AddisGPT
- Location: Virginia, USA