IndicTalk: A Large-Scale Persona-Based Multilingual Conversational Corpus for Indic Languages
Sahil Deepak Gawande, Mayank Singh
cs.CL, cs.LG
2026-07-25
Code-mixed conversational data for Indic languages has been scarce; IndicTalk uses a fully automated news-grounded, persona-conditioned LLM pipeline to generate 1.32M multi-turn conversations across 9 Indic languages in both native-script and Romanized forms.
Bilingual speakers in India have a natural habit: within one conversation they switch back and forth between English and their native language, written either in the native script or in Romanized form. This mixing is called code-mixing. But data to train conversational models on it is essentially absent. Existing Indic NLP datasets are mostly discriminative tasks like sentiment analysis or hate-speech detection; conversational resources are rare and mostly limited to Hinglish at small scale. That is a bottleneck for training and evaluating multilingual conversational LLMs.
IndicTalk fills exactly that gap: a large-scale, event-grounded, multi-turn, code-mixed conversational corpus across many Indic languages and both writing systems.
The pipeline is fully automated. First, raw news and blog documents are summarized by Sarvam-M. Generating dialogue directly from long documents biases the LLM toward verbose output, so the document is first converted into a language-independent semantic representation. Sarvam-M was picked from a blind arena evaluation (Elo 1861, above Llama 7B's 1582 and Gemma 7B's 1056).
Each conversation is conditioned on a pair of personas from five categories: friends, family, colleagues, experts, and student-teacher. The persona stays fixed across the conversation to keep context consistent.
A multilingual LLM (GPT-OSS-120B, which during development passed downstream validation more reliably than Qwen2.5-72B or Gemma 4-31B) generates the dialogue for each language in two variants: native-script (the Indic language in its own script, English in Roman) and Romanized (both languages in Roman script). Each conversation is 6-8 turns.
Finally, automatic validation filters the output. Native-script variants require each utterance to contain at least one target-script token and one Roman token, to meet a minimum code-mixing index (CMI), and to exceed a minimum length. Romanized variants skip CMI because both languages share the Roman script and token-level language identification is unreliable; their quality is checked by human evaluation.
Scale: from about 142K news and blog articles across 12 domains, the pipeline produces over 1.32M conversations, 9 Indic languages times 2 scripts giving 18 varieties, 10.69M turns total, averaging 7.55 turns each. Generation took 16 H200 GPUs and about 10,000 GPU-hours.
Code-mixing metrics land where expected: native-script variants average CMI 37.81, switch-point fraction (SPF) 0.438, and M-index 0.865, indicating frequent switching and roughly balanced bilingual use; Telugu and Kannada mix most strongly. Pseudo-perplexity (PPPL) for native-script variants sits at 9.14-18.52, so the sentences stay fluent despite frequent switching.
| Metric | Native-script | Romanized |
| Avg CMI | 37.81 | 20.23 |
| Avg SPF | 0.438 | 0.311 |
| Avg M-index | 0.865 | 0.508 |
| Pseudo-perplexity (avg) | 13.58 | not reported |
On quality, two judge models (Gemini-2.5-Flash and GPT-OSS-120B, 500 conversations per language, 9,000 total) rate most varieties above 4.0 overall (5-point scale), and their rankings correlate strongly (rho 0.91). In human evaluation (10 per language, 180 total), native-script variants score slightly higher than Romanized; Marathi and Kannada lead native-script, Hindi leads Romanized. Human and LLM rankings correlate only moderately (rho 0.469-0.559).
For teams building conversational systems for low-resource languages, this is a usable training and evaluation corpus, and the fully automated event-grounded, persona-conditioned, validation-gated pipeline is documented clearly enough to reuse for other language families. The caveat is that the value of synthetic conversational data depends on downstream tasks, and the paper does not report an end-to-end gain on any benchmark; it argues the data meets a quality bar.
The authors are direct: this is synthetic data and may miss pragmatic nuance, dialectal variation, and disfluency of real bilingual conversation; sources are news and blogs, so the corpus skews formal and event-centric; human evaluation uses only 2-3 graduate annotators per language; and pseudo-perplexity is reported only for native-script variants.
A structural concern: both generation and judgment lean heavily on GPT-OSS-120B. Using Gemini as an independent judge mitigates self-preference, but having the same model family generate and evaluate the data cannot fully remove the bias. The moderate human-LLM correlation also hints that high LLM scores do not always mean humans find the text natural.