HinglishBench: a parallel English / Hinglish reasoning evaluation for open LLMs

What happens when a 501B reasoning model meets the way a billion people actually talk? Beam-1.0-501B-A23B is "developed predominantly for English" (its own model card). Hinglish - the romanized Hindi-English code-mix used daily across India - is not English. HinglishBench measures the gap.

This is a parallel code-mixed (English/Hinglish) reasoning evaluation for frontier open models: every item exists in two language versions of the same problem, so accuracy and cost differences are attributable to language, not difficulty. (GLUECoS, Khanuja et al. 2020, covered code-mixed encoder tasks; we evaluate generative reasoning.)

Headline results (GSM8K-Hinglish, 82 parallel pairs, zero-shot, single run)

Model	English	Hinglish	Gap (pts)	Median token tax (HI/EN)
Beam-1.0-501B-A23B (reasoning_effort=low)	96.3%	95.1%	-1.2	1.47x
Gemma-3-27B-it	91.5%	93.9%	+2.4	0.96x
Qwen3-32B	90.2%	82.9%	-7.3	1.00x

1. Beam scores highest on Hinglish of the three tested (95.1%) despite English-predominant training - one item ahead of Gemma (93.9%) - but a Hinglish answer costs it a median ~1.5x completion tokens. On these runs, Beam's Hinglish tax shows up as serving cost rather than accuracy.
2. The tax is not universal. Gemma-3-27B shows no Hinglish penalty in these runs (accuracy slightly above English, no token cost). Qwen3-32B scores 7.3 accuracy points lower on Hinglish.
3. The two prompt variants we tested don't remove the tax. A "think in English" system prompt on 30 Hinglish items changed accuracy by 0 points and increased tokens by 9.9%. reasoning_effort=medium does not buy it back either (observed tax 1.39x vs 1.47x; one sample per cell, so we do not estimate run-to-run variance. BeamDial likewise reports effort levels tying on standard math).
4. A second controlled set (100 author-built parallel pairs in easy/medium/hard tiers, Beam only) replicates the tax: 98.0% = 98.0% accuracy, 1.58x median token tax.
5. Reverse-direction replication. A third set (309 parallel pairs: Beam-authored Hinglish-first problems translated to English) shows Beam's tax again - 1.30x - with accuracy saturated (99.4% EN / 99.7% HI). Gemma 0.99x and Qwen 0.94x again show no tax. So the tax is not limited to English-first GSM8K items translated into Hinglish; the reverse set is itself translated, Beam-authored and easier, so it bounds but does not eliminate translation/selection explanations (see Limits).

What's in this dataset

• gsm8k_hinglish.jsonl - 82 parallel pairs: GSM8K test items (English) + romanized-Hinglish translations, with gold answers.
• tiered_100.jsonl - 100 author-built parallel pairs in easy/medium/hard difficulty tiers, with gold answers.
• hinglish_reasoning_beam.jsonl - 326 original Hinglish-first math problems authored by Beam (romanized, Indian context, 12 topics x easy/medium/hard), each with a numeric gold answer verified by self-consistency.
• beamgen_rejected.jsonl - the 7 generation/solve mismatches and automated-check rejects, with reasons.
• beamgen_parallel.jsonl - 309 parallel EN/Hinglish pairs: the reverse-direction set (Hinglish-first items
  • their English translations), used for headline point 5.
• beamgen_parallel_excluded.txt - the 17 translation-check exclusions from that set, with the same numeric-preservation rule as excluded_items.txt.
• predictions/ - per-item raw outputs, parsed predictions, correctness and token usage for every (model, language, item) cell, including the reasoning_effort=medium, mitigation, and reverse-direction runs.

Methodology

• Items: 100 GSM8K test items, uniform random sample (seed 42). Translated to natural spoken-style romanized Hinglish by Beam itself (reasoning_effort=low) - we verified every translation with an automated numeric-token preservation check and excluded 18 items that failed it (IDs listed in excluded_items.txt), leaving 82 pairs. During preparation, 5 kept translations were additionally read end-to-end and found faithful (spot-check performed by the author's AI assistant, not an independent human reviewer).
• Prompting (all models, identical): "Solve the problem. You may think briefly, then end your reply with a final line containing ONLY the numeric answer." Zero-shot, no few-shot. (An earlier answer-only prompt was scrapped: it unfairly suppresses non-reasoning models, dropping Gemma to 25%.)
• Grading: numeric exact match on the final line (fallback: last number in output).
• Inference: HF Inference Providers. Beam via Baseten at reasoning_effort=low (plus a medium run); Qwen3-32B with thinking disabled; Gemma-3-27B-it default. Single sample per cell.
• Token tax: per-item ratio of completion tokens (Hinglish / English), reported as the median over pairs.
• Beam-authored Hinglish set: Beam generated 340 original romanized-Hinglish word problems (12 everyday topics, 3 difficulty tiers, reasoning_effort=low). 333 survived automated checks (Latin script only, >=2 Hinglish marker words, length bounds) plus near-duplicate filtering. A fresh Beam instance then solved each item independently; the 326 items where the solver's answer matched the authored key (numeric exact match) were kept; the 7 disagreements went to beamgen_rejected.jsonl. A random 30-item sample of kept items was re-solved and read during preparation (performed by the author's AI assistant, not an independent human reviewer): 30/30 keys correct, Hinglish natural.
• Reverse-direction set: the 326 verified Hinglish-first items were translated to natural English by Beam (reasoning_effort=low) under the same automated numeric-preservation check (with comma/word-number normalization). 17 items failed (13 empty translations, 2 truncations, 2 word-number renderings; IDs in beamgen_parallel_excluded.txt) and were excluded, leaving 309 pairs. Same prompt, grading, and paired-median token-tax definition as the GSM8K-Hinglish runs; all three models ran under the identical fair prompt used for the headline table.
• Cost: total inference spend for the entire project (all pilots, calibration, translations, every reported run, and this reverse-direction run: ~3,700 generations) measured $0.95 against HF billing on Oct 23, 2026. Per-call token usage is logged in the prediction files.

Limits (read before citing)

• Single run per cell, zero-shot; n=82. We report observed differences without claiming statistical significance (e.g. Beam low vs medium effort differs by about one item).
• Romanized Hinglish only - no Devanagari, no other code-mixed pairs.
• Hinglish translations are model-generated (Beam) and automatically verified, not human-certified gold.
• GSM8K is English-US-context word math; the translation preserves context (dollars stay dollars).
• One provider per model, one temperature (default), one prompt format.
• Author-built tiered items are easier than GSM8K and share one author's style.
• Beam-authored items were both written and verified by Beam (self-consistency, not human gold); a 30-item sample re-checked by the author's AI assistant passed, but residual key errors at the ~1-2% level are plausible.
• The reverse-direction set is still translated (Hinglish to English), Beam-authored, and easier than GSM8K: accuracy saturates there (Beam 99%+), so it evidences the token tax replicating beyond English-first translated items, not a full refutation of translation artifacts; Beam-authored topic/style selection and the 17 excluded items can bias it.

Why this matters

Serving costs scale with output tokens. If Hinglish answers cost ~1.5x English answers at equal accuracy, serving India's code-mixed majority on Beam costs ~1.5x more per answer - invisible in English-only evals. The fix is not prompting (we tested); candidate fixes live in tokenization and training data. Beam's own translations passed 82% automated verification, which suggests open models can already bootstrap code-mixed data.

Citation

@misc{hinglishbench2026,
  title  = {HinglishBench: a parallel English/Hinglish reasoning evaluation for open LLMs},
  author = {Kharya, Sanskar},
  year   = {2026},
  howpublished = {Hugging Face dataset, Build with Beam community challenge entry}
}

GLUECoS: Khanuja et al., "GLUECoS: An Evaluation Benchmark for Code-Switched NLP", ACL 2020. GSM8K: Cobbe et al., "Training Verifiers to Solve Math Word Problems", 2021 (MIT license).