id int64 | category string | english string | code_switched string |
|---|---|---|---|
1 | API Backend & Non-blocking queues | Write a FastAPI endpoint that handles background tasks using a non-blocking queue without blocking the API response time. | FastAPI mein ek aisa endpoint banao jo background tasks ko non-blocking queue ke through handle kare taake API response time block na ho. |
2 | RAG & Chunking Strategy | Explain how section-aware chunking improves retrieval precision in a RAG pipeline. | Yeh batao ke RAG pipeline ke andar section-aware chunking retrieval precision ko kaise behtar karti hai? |
Code-Switched Technical Discourse & Contextual Fragility Evaluation
1. Identification of the Blind Spot
Standard AI benchmarks evaluate models exclusively in formal English. However, in regions like Pakistan, technical collaboration naturally occurs via code-switching (interspersing technical English programming constructs with regional conversational syntax or Roman Urdu). Small frontier models experience severe instruction-following degradation when handling these queries.
2. Evaluation Findings
Tested on meta-llama/Llama-3.2-3B-Instruct:
- English Prompt: Handled with formal architectural tooling recommendations (e.g., Celery, Redis).
- Code-Switched Prompt: Triggered linguistic shifts, recommending lightweight native Python alternatives (
BackgroundTasks,asyncio) and altering structural output formatting.
3. Proposed Path Forward
- Targeted instruction tuning using synthetic code-switched preference datasets.
- Language-invariant intermediate normalization adapters to map regional code-switched inputs into clean semantic intent vectors.
- Downloads last month
- 32