Sachin
@sachdh
cooking reasoning models and agents at @AthenaAgentRL
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
working with @_deepuu was absolute pleasure. he has some crazy ideas + insane resilience to see those come true. if you are interested in building with @physics__wallah , you should ping him
Proud to share Aryabhata 1.0, our first attempt at building India-centric, exam-focused SLMs. Looking forward to feedback. Happy to discuss if you want to contribute in this mission :) @physics__wallah @sachdh @RitvikRastogi19
Brilliant execution by @sachdh!
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Tremendous work
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
feeling vindicated for having shorter context lengths 🤪 jk - that was an allocation of compute decision than anything else
Anthropic just released a research paper. Inverse Scaling in Test-Time Compute This study shows that longer reasoning in Large Reasoning Models (LRMs) can hurt performance—revealing a surprising inverse scaling between reasoning length and accuracy. According to this paper,…
this model looks very fun from initial testing curated a private benchmark today morning, waiting for my laptop to arrive, will test it out!
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Kudos. Awesome. Jai ho.
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Seeing @sachdh at it at RL when it wasn’t as cool since the first time we met to eventually pull this off is testament to the sheer grit and depth needed. Congrats man! Maybe the training go longer! Report please 😋
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
lfg 🔥
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
This is brilliant. Congrats @sachdh. Yes, Indians can cook; And this is done from Bangalore also! Focuses on maths for now!
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Also the model name picked here is classy. If ya folks don't know who is aryabhatta, please read up 😊
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Also the model name picked here is classy. If ya folks don't know who is aryabhatta, please read up 😊
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
grok explaining what we cooked and giving approval
Yes, Aryabhatta 1.0's 90.2% on JEE Mains April 2025 is verified via their Hugging Face repo and benchmarks, outperforming o1-mini and Gemini Flash there. OpenAI's o3/o4 models did score ~327/360 on JEE Advanced mocks (equiv. AIR ~4), but that's a tougher exam—Aryabhatta focuses…
mangalamsx Aryabhatta 1.0 is a 7B parameter AI model by AthenaAgentRL and Physics Wallah, specialized for math in Indian competitive exams like JEE Mains, scoring 90.2%—outperforming models like o1-mini and Gemini Flash. Yes, it's focused on edTech for such exams, trained on…
Amazing work! Congratulations to your entire team! 😀 #athenaAgent
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Pratyush explained this better than I could. We aren’t solving for general intelligence. But on a narrow task, where users care we can deliver similar or slightly better performance at much cheaper cost and latency.
I've followed @sachdh's love for RL for 2+ years at this point & it's so great to see him help ship this This is interesting for a whole host of reasons but in particular because "scope" is beating "scale" Enterprise AI is tilting from “giant‑generalist” AGI aspirations toward…
Incredible feat by @sachdh Aryabhatta-1.0 scores 90.2% on JEE Mains, outperforming o4 mini and Gemini Flash 2.5. If you or your organization is looking for RL-as-a-service solutions, Sachin is the go-to person.
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Power of Indian Tech in AI space.
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
I have seen Sachin bring every ounce of his love for RL and neural net training to bear for this. He is one of the best research scientists in India and the world. RL is the future, watch the early outputs of somebody who is going to be a master in RL.
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…
Can confirm this was trained with lots of reinforcement and hard work. Wildly impressive - go look! ✨
Excited to share Aryabhatta 1.0, our leading model that scores 90.2% on JEE Mains, outperforming frontier models like o4 mini and Gemini Flash 2.5 Trained by us at @AthenaAgentRL , in collaboration with @physics__wallah, using custom RLVR training on 130K+ curated JEE problems…