Hindi
One of the major languages of India, belonging together with Bengali to the Indo-European language family, and frequently used as a related language in cross-lingual transfer for low-resource languages.
Timeline (last 90 days)
Transferring from Hindi to Bengali coreference resolution did not yield consistent gains across models.
Measured with the Qwen3 tokenizer on FLORES devtest sentences, Hindi consumes 4.44 times as many tokens per sentence as English.
For the same Hindi sentence using different tokenizers, the token cost multipliers are: Qwen3/Qwen2.5 4.44 times, Sarvam-M 1.79 times, and Gemma-3 1.31 times.
On Qwen3, romanization (ITRANS transliteration) reduces Hindi token overhead from 4.44x to 2.06x.
On Gemma-3, romanization instead worsens Hindi token overhead from 1.31x to 1.85x.
Many FLORES sentences are not in Unicode NFC normalized form: 609 of 1,012 Bengali sentences, 573 Punjabi sentences, and 93 Hindi sentences; failing to normalize first causes token count drift.
BLEU is sensitive to morphological variation, and Hindi, Urdu and Sindhi all have rich morphological inflection, where the same meaning can have multiple morphological realizations.
Within Hindi or Spanish, the correlation coefficient between a translated item's score and its own rewording score is as high as 0.84 and 0.82, far higher than its cross-language correlation with the original English item.
For Hindi, Korean and Swahili, the score drops brought about by cross-language transfer clearly exceed the fluctuations produced by English rewording.
Most of the cross-language score drops for Hindi, Korean and Swahili come from specific test items in the StereoSet dataset, rather than from a global fairness improvement.
5 more timeline events
All Facts (15)
Transferring from Hindi to Bengali coreference resolution did not yield consistent gains across models.
50%UnverifiedMeasured with the Qwen3 tokenizer on FLORES devtest sentences, Hindi consumes 4.44 times as many tokens per sentence as English.
50%UnverifiedFor the same Hindi sentence using different tokenizers, the token cost multipliers are: Qwen3/Qwen2.5 4.44 times, Sarvam-M 1.79 times, and Gemma-3 1.31 times.
50%UnverifiedOn Qwen3, romanization (ITRANS transliteration) reduces Hindi token overhead from 4.44x to 2.06x.
50%UnverifiedOn Gemma-3, romanization instead worsens Hindi token overhead from 1.31x to 1.85x.
50%UnverifiedMany FLORES sentences are not in Unicode NFC normalized form: 609 of 1,012 Bengali sentences, 573 Punjabi sentences, and 93 Hindi sentences; failing to normalize first causes token count drift.
50%UnverifiedBLEU is sensitive to morphological variation, and Hindi, Urdu and Sindhi all have rich morphological inflection, where the same meaning can have multiple morphological realizations.
50%UnverifiedWithin Hindi or Spanish, the correlation coefficient between a translated item's score and its own rewording score is as high as 0.84 and 0.82, far higher than its cross-language correlation with the original English item.
50%UnverifiedFor Hindi, Korean and Swahili, the score drops brought about by cross-language transfer clearly exceed the fluctuations produced by English rewording.
50%UnverifiedMost of the cross-language score drops for Hindi, Korean and Swahili come from specific test items in the StereoSet dataset, rather than from a global fairness improvement.
50%UnverifiedHindi belongs to the Indo-European language family and retains a fine-grained vocabulary distinguishing patrilateral from matrilateral kin, such as the father's elder brother 'taau' versus the mother's brother 'maama'.
50%UnverifiedHindi has multiple dialects and accent variations, posing challenges for speech recognition models.
50%UnverifiedAchieving equivalent LLM performance quality in a Hindi-language environment requires extensive training and fine-tuning on local language corpora.
50%UnverifiedNative Hindi speakers are estimated to number over 500 million.
50%UnverifiedAlexa+ introduces support for Hindi in India for the first time.
50%