Tamil-ai/tamil-morphological-benchmark

Tamil Morphological Generalization Benchmark (TAMIL-MORPH)

The first morphological generalization benchmark for Tamil -- 1,030 test cases across 9 categories designed to evaluate whether LLMs truly understand Tamil morphological rules or merely memorize surface forms.

Paper: "A Thousand Language Problem: Morphological Understanding in Linguistic AI"

Benchmark Overview

Category	Test Cases	Description
Case Suffixes (வேற்றுமை)	240	6 grammatical cases across 40 noun roots
Plural + Case (பன்மை)	~160	Plural formation with case markers
Verb Conjugation (வினைத்திரிபு)	~210	7 person-tense combinations across verb roots
Sandhi (புணர்ச்சி)	~50	Sound changes at word boundaries
Honorific Forms (மரியாதை)	~90	Informal/formal/high-respect registers
Negation (எதிர்மறை)	~90	Present/past/future negative forms
Compound Words (கூட்டுச்சொல்)	~50	Word joining rules
Conditional/Causal (நிபந்தனை)	~60	Conditional and causal suffixes
Novel Combinations (புதிய வடிவங்கள்)	~80	Multi-suffix forms never seen in training
Total	1,030

Baseline Results

Model	Overall Accuracy
GPT-4o-mini	54.0%

Files

• Benchmarkdata.md -- Full benchmark data (JSON arrays in Markdown)
• morph_benchmark_eval.py -- Complete evaluation script (supports local HF models, OpenAI, Google Gemini backends)
• baselines/gpt-4o-mini_results.json -- Detailed per-test results for GPT-4o-mini
• kaggle_benchmark.ipynb -- Ready-to-run Kaggle notebook for benchmarking
• runpod_benchmark.py -- RunPod GPU benchmarking script

Usage

Run evaluation locally

# With OpenAI API
python morph_benchmark_eval.py --model gpt-4o-mini --backend openai

# With Google Gemini (free tier)
python morph_benchmark_eval.py --model gemini-2.0-flash --backend gemini

# With local HuggingFace model
python morph_benchmark_eval.py --model Tamil-ai/tamil-qwen25-7b-instruct --backend local

# Run all configured models
python morph_benchmark_eval.py --all

Load benchmark data programmatically

from huggingface_hub import hf_hub_download
import json, re
from pathlib import Path

path = hf_hub_download(
    repo_id="Tamil-ai/tamil-morphological-benchmark",
    filename="Benchmarkdata.md",
    repo_type="dataset",
)
# Parse JSON blocks from the markdown (see morph_benchmark_eval.py for full parser)

Data Format

Each category contains structured JSON with roots, meanings, and expected morphological forms:

{
  "root": "வீடு",
  "root_meaning": "house",
  "forms": {
    "accusative": {"tamil": "வீட்டை", "meaning": "the house (object)"},
    "dative": {"tamil": "வீட்டுக்கு", "meaning": "to the house"},
    "locative": {"tamil": "வீட்டில்", "meaning": "in the house"}
  }
}

Scoring

• 1.0 -- Exact match (after Tamil text normalization)
• 0.5 -- Partial match (predicted is substring of expected)
• 0.0 -- Wrong

Why This Benchmark?

Existing Tamil NLP benchmarks test translation or classification. None test whether models understand the generative morphological rules of Tamil -- an agglutinative language where a single root can produce hundreds of valid surface forms through suffix combinations.

This benchmark is transferable to other agglutinative languages (Turkish, Finnish, Hungarian, Korean, etc.) by replacing the morphological rules.

Validation

All 1,030 test cases were validated using:

1. Finite State Transducer (FST) analysis
2. Stanza NLP morphological parser
3. Manual rule verification

Citation

@misc{tamilmorph2026,
  title={A Thousand Language Problem: Morphological Understanding in Linguistic AI},
  author={Tamil-AI},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/Tamil-ai/tamil-morphological-benchmark}
}