O96a/sudanese-synthetic-instructions

Synthetic Sudanese Arabic Instruction Data Generator

Experiment exp-004 | Sudanese NLP domain (PRIORITY) | April 12, 2026

Overview

This experiment tests structured synthetic data generation for Sudanese Arabic instruction following, inspired by the "Agent-as-Annotators" paper's approach to structured trajectory synthesis.

Hypothesis

Structured synthetic dialogue generation using seed patterns can create viable instruction-following data for Sudanese Arabic, even with minimal authentic examples.

Method

1. 1.Seed with authentic Sudanese Arabic conversational patterns
2. 2.Apply template-based variation
3. 3.Generate instruction-response pairs across categories
4. 4.Compare with MSA equivalents

Categories Covered

• —Greetings: "هاي كيفك؟", "صباح الخير", etc.
• —Daily Tasks: "شنو بتعمل النهارده؟", "وين رايح؟"
• —Requests: "ممكن تجيب لي مي؟", "عندك فكة؟"
• —Food: "شنو الغدا؟", "جعان كتير"

Key Findings

Sudanese vs MSA Differences

Synthetic Data Quality

• —Lexical diversity: Limited by seed pattern diversity
• —Dialect authenticity: High for common phrases, needs expansion
• —Scalability: Template-based approach scales linearly
• —Validation needed: Requires native speaker evaluation

Applications

1. 1.Bootstrap fine-tuning for Sudanese chatbots
2. 2.Create evaluation benchmarks
3. 3.Study dialect-specific instruction following
4. 4.Augment limited authentic datasets

Files

• —app.py: Gradio interface for data generation
• —requirements.txt: Pinned dependencies
• —README.md: This file

Next Steps

• —[ ] Collect real Sudanese conversations for validation
• —[ ] Fine-tune small LLM (1B-3B) on synthetic data
• —[ ] Evaluate against native speaker judgments
• —[ ] Expand pattern library with more categories

References

• —Agent-as-Annotators paper: https://huggingface.co/papers/2604.07776
• —Related: exp-003 (Sudanese Arabic OCR Stress Test)

Space

https://huggingface.co/spaces/O96a/sudanese-synthetic-instructions

Build trigger

Rebuild trigger Sun Apr 12 18:02:07 UTC 2026

Sun Apr 12 20:40:20 UTC 2026