README.md

metadata

license: cc-by-4.0
language:
  - ta
  - en
pretty_name: TamilBench
size_categories:
  - n<1K
task_categories:
  - question-answering
tags:
  - tamil
  - tanglish
  - code-mixed
  - romanized
  - beam
  - build-with-beam
  - token-efficiency
configs:
  - config_name: parallel
    data_files: data/parallel.jsonl
  - config_name: native
    data_files: data/native.jsonl
  - config_name: beam_runs
    data_files: results/beam_runs.jsonl

TamilBench v1

How well does Beam reason in Tamil, and what does it cost?

• parallel (50 items): the same question in English, Tamil script and colloquial Tanglish (romanized Tamil). Categories: math word problems with Indian context (20), logic (8), Tamil culture/knowledge (15), everyday (7). Short gold answers plus aliases.
• native (22 items): questions that only exist in Tamil: Tamil number words, ancient Tamil numerals (௲ ௱ ௰), letter counting by Tamil grammar (எழுத்து), Sangam/Thirukkural literature, grammar, the Tamil calendar.
• beam_runs (516 rows): every response from reflection/Beam-1.0-501B-A23B (HF Inference Providers → Baseten) with output, reasoning trace, prompt/reasoning/completion tokens, latency, correctness and cost.

Headline results (Beam-1.0-501B-A23B, Oct 2026)

Language	Accuracy	Prompt tok	Reasoning tok	$ per 1k queries	$ per 1k at reasoning_effort=minimal
English	100%	90	106	0.178	0.150
Tamil script	100%	231	161	0.301	0.254
Tanglish	100%	104	214	0.313	0.246

• Beam's tokenizer spends 8.0 tokens per Tamil word (0.92 chars/token) vs 1.5 for English: the same question costs 3.75× more input tokens in Tamil script.
• Tanglish is cheap to read (2.4 tokens/word) but Beam thinks 2× longer (4× on logic puzzles): it translates internally first.
• Net: Tamil and Tanglish queries cost ~1.7× English. reasoning_effort="minimal" kept 100% accuracy and cut cost 16–21%.
• Native Tamil split: 98.5% (65/66); the one miss was the last word of Kural 1.

Scoring: last number for numeric answers; alias match for words. Prices: Baseten list $0.40/1M input, $1.20/1M output (reasoning billed as output). The parallel set is saturated for Beam, so read accuracy as a sanity floor; the token/cost findings are the contribution.

Interactive explorer + live Beam tokenizer: https://huggingface.co/spaces/esscor/TamilBench

Built by Mohamed Adib for the Hugging Face × Reflection Build with Beam challenge.