Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringclasses
6 values
category
stringclasses
2 values
prompt
stringclasses
6 values
expected
stringclasses
6 values
model_output
stringclasses
6 values
geez_add_01
geez_arithmetic
Evaluate this Ge'ez numeral expression and write the answer in both Ge'ez and Arabic numerals: ፭ + ፯ = ?
12 (፲፪)
The Ge'ez numeral expression ፭ + ፯ translates to 7 + 8 in Arabic numerals. Therefore, ፭ + ፯ = 15 in both Ge'ez and Arabic numerals.
geez_mul_02
geez_arithmetic
What is ፳ multiplied by ፫? Provide the result in Ge'ez numerals.
60 (፷)
The product of ፳ (3) and ፫ (45) is 135, which in Ge'ez numerals is ፻፰ (\u16e9).
geez_sub_03
geez_arithmetic
Compute: ፻ - ፳፭ = ?
75 (፸፭)
The given expression seems to be in an unusual numeral system or notation that is not standard (likely using Ethiopic numerals). However, assuming the first number is ፻ (which could represent 10 in some systems) and the second number is ፳፭ (which could represent 37 in some systems), we can perform the subtraction. If ...
calendar_modular_01
ethiopian_calendar
በኢትዮጵያ ዘመን አቆጣጠር መስከረም 1 ቀን 2016 ዓ.ም ከሆነ፣ ከ 45 ቀናት በኋላ ምን ቀን እና ወር ይሆናል? (በኢትዮጵያ የቀን መቁጠሪያ እያንዳንዱ ወር 30 ቀናት አሉት)
ጥቅምት 16 ቀን 2016 ዓ.ም
የኢትዮጵያ የቀን መቁጠሪያ በአጠናught ወር 30 ቀናት አሉት። ከሆነ በዓመት 2016 ዓ.ም፣ ከሆነ 45 ቀናት በኋላ ይሆናል። 45 ÷ 30 = 1.5 1.5 ቀናት ይሆናል። ከሆነ 1 ቀናት ይሆናል፣ ከሆነ 0.5 ቀናት ይሆናል። በአጠናught ወር ከሆነ 1 ቀናት ይሆናል፣ ከሆነ 0.5 ቀናት ይሆናል። ከሆነ 0.5 ቀናት
calendar_pagume_02
ethiopian_calendar
በኢትዮጵያ የቀን መቁጠሪያ መሰረት ከነሐሴ 28 ቀን በኋላ 7 ቀናት ቢጨመሩ የትኛው ቀን እና ወር ይሆናል? የጳጉሜን 5 ቀናት አስብ።
ጳጉሜ 5 (ወይም መደበኛ ዓመት ከሆነ መስከረም 1)
የኢትዮጵያ የቀን መቁጠሪያ መሰረት ከነሐሴ 28 ቀን በኋላ 7 ቀናት ቢጨመሩ የትኛው ቀን እና ወር ይሆናል። በአጋግናና የጳጉሜን 5 ቀናት አስብ ነው። የተደጋጋሚ ግራঢ়ঞ্চ ነው ግራঢ়ঞ্চ በአጋግናና የጳጉሜ የተደጋጋሚ ቀን እና ወር ይሆናል። በአጋግናና የጳጉሜ የተደጋጋሚ ቀን 5 ቀናት አስ
calendar_year_gap_03
ethiopian_calendar
የኢትዮጵያ ዘመን አቆጣጠር ከጎርጎሪዮሳዊ (Gregorian) አቆጣጠር በስንት ዓመታት ይለያል? ምክንያቱን በአጭሩ ግለጽ።
7 or 8 years difference (due to the Annunciation calculation)
የኢትዮጵያ ዘመን አቆጣጠር እናገኙ Gregorian አቆጣጠር 1974-09-12 ይለያል. በአጭሩ ግለጽ ነው 1974-09-12.

Blind Spots of Frontier Models: Tokenizer-Induced Numerical & Temporal Collapse in Ge'ez and Amharic

  • Author: Teshome Birhanu Cheru
  • Affiliation: Addis Ababa University, Electrical and Computer Engineering
  • Target Fellowship: Fatima Fellowship 2026 Technical Challenge
  • Evaluated Model: Qwen/Qwen2.5-3B-Instruct (3B Parameters)
  • Artifacts Repository: Evaluation Benchmark (benchmark_ethiopian_reasoning.json), Empirical Outputs (eval_results.json), Colab Notebook (qwen_ethiopian_reasoning_eval.ipynb)

1. Identified Blind Spot & Motivation (Lived Experience)

Growing up in Ethiopia and completing my engineering studies at Addis Ababa University, daily civic, cultural, and logistical activities exist under a dual cognitive framework: the Western Gregorian system alongside the traditional Ethiopian (Ge'ez) calendar and numeral heritage.

The Ethiopian calendar operates on a distinct astronomical and computational structure: twelve uniform months of 30 days, followed by a thirteenth transitional month (Pagumē / ጳጉሜ) consisting of 5 days (6 in leap years), lagging roughly 7 to 8 years behind the Gregorian calendar. Concurrently, Ge'ez numerals (፩, ፪, ፫ ... ፲, ፳, ፻) represent an additive-decimal system without a zero symbol, historically and actively embedded in religious records, formal certificates, and national documents.

Standard NLP benchmarks (MMLU, GSM8K, MATH) evaluate models almost exclusively on English and high-resource Latin-script tokens. The critical capability gap identified here is:

Current frontier open-weight models suffer severe sub-word and byte-level tokenizer fragmentation over Ethiopic scripts, causing an immediate breakdown in additive numeral arithmetic, symbol recognition, and non-Gregorian modular calendar logic.

While modern open-weight models have achieved near-perfect accuracy on grade-school arithmetic in English, translating the identical mathematical problem into Ge'ez numerals or Amharic calendar constraints exposes two failure modes:

  1. Sub-word Fragmentation: The Byte-Pair Encoding (BPE) tokenizer fragments individual Ge'ez numerals into disjoint byte chunks, obscuring the numerical values from the model's internal representations.
  2. Cultural & Ontological Bias: Models default to Western Gregorian assumptions (e.g., 31-day months, omitting Pagumē), producing confident hallucinations or decaying into repetitive degenerate tokens.

2. Systematic Evaluation & Empirical Findings

I systematically stress-tested Qwen/Qwen2.5-3B-Instruct (a leading 3B open-weight model with claimed state-of-the-art multilingual and mathematical capabilities) using a free T4 GPU environment in Google Colab. The evaluation used deterministic greedy decoding (do_sample=False) over paired tests across three categories.

A. Empirical Proof of Tokenizer Fragmentation

Direct inspection of the tokenizer vocabulary revealed severe byte-level fragmentation for Ethiopic glyphs:

Input Glyph Meaning Tokenized Representation Token Count English Equivalent
፭ 5 ['áį', 'Ń'] 2 sub-byte tokens 5 (1 token)
፲ 10 ['áį', '²'] 2 sub-byte tokens 10 (1 token)
፳፭ 25 ['áį', '³', 'áį', 'Ń'] 4 sub-byte tokens 25 (1 token)
ጳጉሜ Pagumē (Month 13) ['áĮ', '³', 'áĮ', 'ī', 'áĪľ'] 5 sub-byte tokens Month (1 token)

While a Latin digit like 25 is encoded as a single token, the equivalent Ge'ez numeral ፳፭ is split into 4 disjoint byte chunks. The model never observes the symbol as a unified semantic or mathematical entity.

B. Evaluation Results on Arithmetic & Temporal Tasks

Test ID Category Query / Problem Expected Truth Model Output Summary Result
geez_add_01 Numeral Arithmetic ፭ + ፯ = ? 12 (፲፪) Translates ፭ + ፯ to 7 + 8, outputs 15 Failed (0%)
geez_mul_02 Numeral Arithmetic ፳ × ፫ = ? 60 (፷) Asserts ፳ = 3 and ፫ = 45, computes 135 (፻፰) Failed (0%)
geez_sub_03 Numeral Arithmetic ፻ - ፳፭ = ? 75 (፸፭) Assumes ፻ = 10 and ፳፭ = 37, computes 10 - 37 = -27 Failed (0%)
calendar_modular_01 Modular Calendar Date 45 days after Meskerem 1 Tikimt 16 Computes 45 / 30 = 1.5, hallucinates 1.5 ቀናት ይሆናል። Failed (0%)
calendar_pagume_02 Leap-Month Logic 7 days after Nehase 28 (via Pagumē) Pagumē 5 / Meskerem 1 Text degeneration; outputs non-Ethiopic script (ግራঢ়ঞ্চ) Failed (0%)
calendar_year_gap_03 Temporal Offset Year offset from Gregorian 7–8 Years (Incarnation) Outputs static corrupted date string: 1974-09-12 Failed (0%)

C. English Control Baseline

When the identical mathematical operations were evaluated in English using Arabic numerals (e.g., 5 + 7, 20 * 3, 100 - 25), the model scored 100% accuracy. This contrast demonstrates that the failure is not caused by a deficit in mathematical reasoning capacity, but rather by severe tokenizer representation blind spots and cultural pretraining biases.


3. Proposed Path Forward

To resolve this capability gap without requiring massive, computationally prohibitive foundation model pretraining from scratch, I propose a three-stage intervention:

1. Dedicated Script Tokenizer Adaptation

  • Vocabulary Expansion: Explicitly allocate dedicated tokens in the tokenizer vocabulary for the base Ge'ez numerals (1–10, 20, 30, ... 100, 10,000) and frequent Amharic morphological roots rather than falling back to multi-byte UTF-8 sequences.
  • Embedding Warm-Start: Initialize the new input and output embeddings as the mean of their constituent byte-level representations to prevent representation shock and preserve pre-existing semantic associations during downstream tuning.

2. Synthetic Isomorphic Reasoning Alignment (Data Curation)

  • Parallel Math Translation: Programmatically generate parallel synthetic reasoning pairs by mapping existing GSM8K and synthetic arithmetic datasets into Ge'ez numerals and Amharic word problems.
  • Chain-of-Thought (CoT) Deconstruction: Fine-tune models with explicit symbolic deconstruction paths: $$\text{Ge'ez Numeral (፳፭)} \longrightarrow \text{Decomposition } (20 + 5) \longrightarrow \text{Compute in Base-10} \longrightarrow \text{Re-encode (፸፭)}$$ This enables the model's existing reasoning heads to operate over familiar integer representations before generating the final script tokens.

3. Calendar-Constrained Multi-Task Fine-Tuning

  • Non-Gregorian Temporal QA: Curate an instruction-tuning corpus incorporating modular date arithmetic that explicitly enforces uniform 30-day boundaries and Pagumē leap transitions.
  • Contextual Switching: Train the model to condition its calendar arithmetic on explicit contextual framing (e.g., distinguishing between Ethiopian fiscal/civic dates and Gregorian dates), mitigating default cultural hallucinations.

Repository Structure

├── README.md                              # Research context, evaluation results, and proposed solutions
├── benchmark_ethiopian_reasoning.json     # Curated benchmark dataset of paired prompts
├── eval_results.json                      # Empirical model outputs from Qwen2.5-3B-Instruct
└── fatima_blindspot_ethiopian_reasoning.ipynb # Reproducible execution notebook (Google Colab / T4 GPU)
Downloads last month
41