BarshaKC/nepali-romanized-evaluation

Evaluating Romanized Nepali and Nepali-English Code-Switching in Qwen2.5-3B-Instruct

Blind Spot

As a Nepali speaker, I have noticed that many people communicate online using Romanized Nepali mixed with English words. This style of communication is common in messaging apps, social media platforms, and student communities. However, most multilingual benchmarks focus on standardized written languages and rarely evaluate informal Romanized language.

As a result, benchmark performance may not accurately reflect how useful these models are for real-world Nepali users. I hypothesized that a modern language model would struggle with Romanized Nepali and Nepali-English code-switched communication despite performing well on standard multilingual benchmarks.

---

Model

Qwen2.5-3B-Instruct

I selected Qwen2.5-3B-Instruct because it is a recent open-weight language model that is large enough to exhibit capabilities similar to current frontier models while remaining small enough to run on Google Colab.

---

Dataset

I manually created a dataset of 23 Romanized Nepali and Nepali-English code-switched sentences based on everyday conversations, social media usage, and student communication.

The dataset includes:

• Common conversational expressions
• Temporal references
• Nepali-English code-switched sentences
• Everyday vocabulary and activities
• Cultural references
• Location references

---

Evaluation Method

The model was asked to translate each Romanized Nepali sentence into English.

Each output was manually scored as follows:

• 2 = Correct translation
• 1 = Partially correct translation
• 0 = Incorrect translation

I also recorded the primary error type for each example.

---

Results

Overall Performance

• Total Examples: 23
• Correct: 0
• Partial: 5
• Incorrect: 18
• Average Score: 0.22 / 2

Common Error Categories

• Tense and aspect errors
• Opposite meaning generation
• Failure to understand common Nepali words
• Hallucinated information
• Failed translations
• Code-switching errors
• Temporal reasoning failures
• Question structure failures

Example Failures

Example 1

• Input: mero exam sakiyo

• Expected Translation: My exam is over

• Model Output: My exam is starting

• Error Type: Opposite meaning generation

Example 2

• Input: ma sathi sanga vetna jadai xu

• Expected Translation: I am going to meet my friend

• Model Output: I will meet you at the station

• Error Type: Hallucinated information

Findings

The model often recognized individual keywords such as exam, assignment, meeting, office, and home. However, it frequently failed to preserve the actual meaning of the sentence.

Several patterns appeared repeatedly:

The model struggled with tense and temporal expressions such as voli, ahile, and sakiyo. Common Romanized Nepali words such as malai, maile, timro, and alxi were often misunderstood. Nepali-English code-switched sentences caused frequent translation errors. The model occasionally hallucinated information that was not present in the input. Outputs were often fluent English sentences but semantically incorrect. These observations suggest that the model has partial word-level understanding of Romanized Nepali but lacks reliable sentence-level understanding.

Proposed Solution

A potential solution is to create larger datasets containing:

1. Romanized Nepali conversations
2. Nepali-English code-switched text
3. Multiple spelling variations of Romanized Nepali words
4. These datasets could be used for continued pretraining or instruction tuning. Additionally, dedicated evaluation benchmarks for Romanized Nepali would help measure real-world performance more accurately than existing multilingual benchmarks.

Conclusion

This evaluation suggests that Romanized Nepali remains a significant blind spot for modern small language models. Although Qwen2.5-3B-Instruct often recognized individual keywords, it struggled to preserve sentence-level meaning and frequently failed on common conversational expressions used by Nepali speakers. The results indicate that performance on standardized multilingual benchmarks may substantially overestimate real-world effectiveness for users who communicate using Romanized Nepali and Nepali-English code-switching.