Post-training data · Reward signals · RL
Most of India speaks Hindi, Hinglish and a dozen other languages that LLMs still handle poorly, because the post-training data does not exist. Bhabha AI builds it, layer by layer: instruction data for SFT, scored data for reward models, verifiable tasks for RL, and the open-source tooling to generate more.
Source: Hugging Face & GitHub public stats
Post-training & RL
Capabilities come from post-training: supervised data first, then preference and reward signals, then reinforcement learning against tasks that can be checked. For Indic languages, almost none of that data exists. We build each layer and release it openly.
10 से विभाजित संख्या 6 है। योंगी ने एक निश्चित संख्या से 15 घटाकर परिणाम प्राप्त किया। उसे क्या परिणाम मिला?
gold answer 45 · checked by exact match
Instruction and chat data in Hindi, Hinglish and English to teach the base behaviour.
1.1M+ rows
Instructions and responses scored for complexity and quality, ready for training reward models and data filters.
160K scored
Problems with checkable answers (math word problems with worked solutions, RAG questions with gold answers) to seed RL with verifiable rewards.
200K+ tasks
Custom environments, graders and rollouts for Indic-language tasks, built with teams post-training their own models.
Custom builds
How we work
Translation alone produces noisy data. Our edge is everything around it: deciding what to translate, catching what breaks, and measuring what is worth training on.
We take the strongest open English instruction, math and RAG corpora and translate them into Hindi and Hinglish with IndicTrans2 — skipping code, formulas and names that must stay untouched.
Every row is checked against its English source. We drop prompts that break under translation (code, wordplay, spelling tasks) and degenerate, repetitive output, then score what remains for quality and complexity, DEITA-style.
Where no source exists, we generate native Hindi data synthetically with Dataformer, our open-source engine for research-backed synthetic data at scale.
We fine-tune models on the result, like Gajendra, then score their outputs and put them in front of real users. What fails becomes the next round of data.
Open source
Open-source library · Python
A framework for creating high-quality synthetic datasets. One async API across OpenAI, Groq, Together, DeepInfra and OpenRouter, with rate-limit handling, response caching and implementations of published data-generation methods.
Model · 7B
Our Hindi-Hinglish-English instruction-tuned chat model, released openly on Hugging Face and quantized by the community for local inference.
Open-source app
A chat interface for serving Indic LLMs side by side, built to collect real-world, openly licensed conversation data and to benchmark Indic models arena-style.
Open datasets
SFT data, scored data for reward models, verifiable math and retrieval tasks, and summarization, in Hindi, Hinglish and English. Used by researchers and builders training Indic models.
English news articles paired with Hindi summaries, translated with IndicTrans2 and de-duplicated. Built for cross-lingual summarization.
About 600K rows of the OpenHermes-2.5 instruction set, translated to Hindi and quality-filtered.
Filtered English–Hindi instruction data derived from AI4Bharat’s indic-instruct, with low-quality and broken translations removed.
First filtered release of indic-instruct; superseded by v0.2 but kept for reproducibility.
Grade-school math word problems with worked solutions, translated to Hindi and filtered.
OpenHermes conversations rewritten into Hindi or Hinglish, for models that must switch scripts and registers.
Hindi passage–question–answer triples, with Hinglish versions of every question, for retrieval-augmented generation.
English contexts with questions in English, Hindi and Hinglish, so models learn to answer across languages.
Hindi translation of Alpaca-GPT4 using IndicTrans2, filtered for translation quality.
Filtered Hindi/English multi-turn conversations for chat fine-tuning.
Synthetic Hindi instructions and long-form answers, generated natively in Hindi from seed topics.
English BBC articles from XL-Sum paired with Hindi summaries for cross-lingual summarization.
Instructions scored for complexity, following the DEITA data-selection method. Used to train third-party complexity scorers.
Instruction–response pairs scored for answer quality, for training data filters.
Labels for whether a prompt should be translated or kept as-is (code, names, math). Powers our translation pipeline.
An improved subset of the Aksharantar Hindi transliteration corpus (Roman ↔ Devanagari).
Models
A 7B instruction-tuned chat model for Hindi, Hinglish and English, built on top of Sarvam AI’s OpenHathi base model.
Decides which prompts are safe to machine-translate and which must stay in the source language, so translated datasets keep code, math and names intact.
Sample output · Gajendra-v0.1
User: Explain the concept of ‘Photosynthesis’ in Hindi and also write the equation.
Gajendra: प्रकाश संश्लेषण एक ऐसी प्रक्रिया है जिसके द्वारा पौधे, शैवाल और कुछ बैक्टीरिया सूर्य के प्रकाश, कार्बन डाइऑक्साइड और पानी का उपयोग करके ग्लूकोज और ऑक्सीजन का उत्पादन करते हैं।6 CO₂ + 6 H₂O + प्रकाश ऊर्जा → C₆H₁₂O₆ + 6 O₂
In the press
Work with us
We build custom SFT and preference data, reward signals, and RL environments with verifiable rewards for teams post-training LLMs for Indian users.