Polite on the Surface, Wrong in Practice: A Curated Dataset for Fixing Honorific Failures in Multilingual Bangla Generation

2026-05-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A new culturally aligned instruction-tuning dataset, BLADE (BangLa Application and DialoguE generation), has been introduced to address critical pragmatic gaps and honorific failures in Multilingual Large Language Models (MLLMs) when generating Bangla text. Comprising 4,196 meticulously curated interaction pairs, this resource facilitates systematic fine-tuning and evaluation of leading open-weight architectures like DeepSeek-8B and LLaMA-3.2-3B. Researchers utilized parameter-efficient fine-tuning via LoRA adapters within a 4-bit NormalFloat (NF4) quantization framework. Empirical evaluations demonstrate that models fine-tuned on BLADE achieve substantial improvements in structural fidelity and honorific alignment, establishing a rigorous benchmark for bridging pragmatic disparities in low-resource multilingual text generation.

Key takeaway

For NLP engineers developing multilingual LLMs for low-resource languages like Bangla, you should prioritize culturally aligned instruction-tuning to mitigate pragmatic gaps. The BLADE dataset offers a proven approach to fix honorific failures and enhance structural fidelity in generated text. Integrate this type of curated resource into your fine-tuning pipeline, potentially using LoRA with 4-bit NF4 quantization, to ensure your models produce culturally nuanced and context-appropriate communication.

Key insights

Culturally aligned instruction-tuning with BLADE dataset significantly improves honorific consistency in multilingual Bangla LLM generation.

Principles

• MLLMs exhibit severe pragmatic gaps in low-resource contexts like Bangla.
• Culturally aligned instruction-tuning enhances structural fidelity and honorific alignment.

Method

Curate 4,196 interaction pairs for Bangla. Fine-tune open-weight MLLMs (e.g., DeepSeek-8B, LLaMA-3.2-3B) using LoRA adapters with 4-bit NF4 quantization.

In practice

• Utilize the BLADE dataset for instruction-tuning MLLMs targeting Bangla.
• Apply LoRA and 4-bit NF4 quantization for efficient fine-tuning of large models.

Topics

• Multilingual LLMs
• Bangla Language Generation
• Honorifics
• Instruction Tuning
• LoRA Fine-tuning
• Dataset Curation

Code references

• ashuvo25/Bangla_Application_LLM

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific datasets and LoRA fine-tuning significantly enhance LLM performance in specialized tasks like classical poetry analysis.
• Cross-Lingual Consensus: Aligning Multilingual Cultural Knowledge via Multilingual Self-Consistency — LLMs can self-align cultural knowledge across languages by identifying consistent responses and transferring them to weaker language contexts.
• Can AI Save India’s Dying Languages? | Building a Tulu LLM Without Training Data Ft. Prathamesh — Structured prompting and custom romanization enable LLMs to generate low-resource languages without extensive training data.
• CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters — CuMA uses demographic-aware routing and specialized adapters to prevent LLMs from collapsing diverse cultural values into a generic average.
• A NATO-accredited organization published a report concerning language-based disparities in LLM performance (Published April 25, 2026) — LLM performance varies significantly by language, but adaptation strategies can bridge the gap.
• LLMs Infer Cultural Context but Fail to Apply It When Responding — LLMs infer cultural context but often fail to apply it in responses without explicit sequential prompting, revealing cultural biases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.