Hinglish-Bench

📄 Paper: Hinglish-Bench — Reference-Free Benchmark for LLM Hinglish Text Generation (gist preprint)

A reference-free benchmark for measuring how well LLMs generate natural Roman-script Hinglish — the Hindi-English code-mixing that hundreds of millions of Indians actually speak, type, and read online.

Reference-free by design. Hinglish has no canonical spelling and no single "correct" rendering, so there are no gold references and no BLEU. Quality is scored by reference-free code-mixing signals plus an LLM-as-judge rubric — the way code-mixed text should be judged.

Why Hinglish-Bench?

1. First of its kind — no established benchmark measures LLM generation of Hinglish prose. Existing resources are training data (translation, ASR, instruction-tuning) or Hindi-only suites that miss code-mixing, Hinglish's defining trait.
2. Realistic registers — 34 prompts across 8 categories of everyday Indian communication: WhatsApp chat, social media, product reviews, storytelling, advice, customer support, captions, and news explainers.
3. Honest metrics — CMI, M-Index, and SyMCoM measure the mixing itself, and a six-dimension judge rubric rates how good the result reads. No misleading BLEU/ROUGE on code-mixed text (HinGE showed they fail).
4. Reproducible — one command generates responses, scores them, and writes a report; ships as a standard lm-evaluation-harness task with CMI, M-Index, and SyMCoM as native metrics.

Dataset Composition

One split (test) with 34 prompts across 8 everyday categories:

Category	Prompts	Share
Casual Conversation	5	14.7%
Social Media Posts	5	14.7%
Advice & Opinions	5	14.7%
Product Reviews	4	11.8%
Storytelling	4	11.8%
Captions & One-liners	4	11.8%
News Explainers	4	11.8%
Customer Support Dialogues	3	8.8%
Total	34	100%

Data Fields

Column	Type	Description
id	string	Unique prompt id (e.g. casual_chat_001)
category	string	Category key (e.g. casual_chat)
category_name	string	Human-readable category name
prompt	string	The generation prompt (elicits natural Hinglish)
expected_style	string	Rubric hint for the desired register/tone
language	string	Always hinglish (code-mixed hi/en)
split	string	Always test

Usage

from datasets import load_dataset

ds = load_dataset("saidutta69/hinglish-bench")
print(f"Prompts: {len(ds['test'])}")
print(ds["test"][0]["prompt"])

# Filter by category
chat = ds["test"].filter(lambda x: x["category"] == "casual_chat")
news = ds["test"].filter(lambda x: x["category"] == "news_explainer")

How to Evaluate

Run the reference implementation (in the project's hinglish_bench package) against any litellm-supported model:

pip install -e .        # installs litellm
hinglish-bench run --model openai/gpt-4o-mini --judge-model openai/gpt-4o

Every response is scored two ways:

1. Automatic code-mixing signals (reference-free):
   • CMI (Gambäck & Das 2014) — how much mixing occurs
   • M-Index (Barnett et al. 2000) — balance across languages
   • SyMCoM (Kodali et al., Findings of ACL 2022) — syntactic evenness of mixing
2. LLM-as-judge rubric (1–10 each): fluency, code-mixing naturalness, Hindi grammar, prompt adherence, spelling consistency, overall.

The benchmark also ships as a standard lm-evaluation-harness task (hinglish_bench.yaml) reporting CMI, M-Index, and SyMCoM natively, plus a documented protocol for validating the judge against human raters (Pearson / Spearman correlation and Fleiss' κ).

The Honest Tradeoffs

• 34 prompts is a starter set — coverage of dialects, formal registers, and Devanagari-script Hinglish is future work. Contributions (new categories, harder prompts, gold references for agreement studies) are welcome.
• LID bias — automatic signals rely on word-level language identification; unknown Roman-script Hindi words default to English, slightly under-reporting Hindi share. Consistent across models, so relative rankings stay meaningful.
• Judge bias — LLM judges can prefer verbose or English-flavoured output; use a strong judge at temperature 0 and always report the judge model.

Citation

@misc{hinglishbench,
  author = {RACER IS OP},
  title = {Hinglish-Bench -- Reference-Free Benchmark for LLM Hinglish Text Generation},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saidutta69/hinglish-bench}}
}

Licensed under CC-BY-SA-4.0.