Suyash Shringarpure retweetledi
Suyash Shringarpure
125 posts
Suyash Shringarpure
@suyashss
Scientist at 23andMe. Interests: Machine learning, Genetics, Privacy
Katılım Eylül 2008
336 Takip Edilen153 Takipçiler
Suyash Shringarpure retweetledi
Delighted to share our latest research from the @23andMeResearch Team, just published in @Nature
! We looked at data from >27,000 participants to uncover how human genetics influences weight loss efficacy and side effects of GLP-1 medications like semaglutide. A thread 🧵👇
English
Suyash Shringarpure retweetledi
Delighted to see our method, PRSformer, at #NeurIPS2025! PRSformer is AI model for population-scale disease-risk prediction from individual genomes. It lays the groundwork for phenome-wide risk prediction. biorxiv.org/content/10.110…
English
Suyash Shringarpure retweetledi
🚨New work led by #CZBiohubCHI Investigator Aly Khan & @23andMeResearch accepted to #NeurIPS2025. PRSformer: disease risk prediction from genetic variants at population scale! New scaling law (N>1M) + multitask + multi-ancestry gains for 18 immune traits.
biorxiv.org/content/10.110…
English
@kalomaze @HrishbhDalal Similar experience trying gene ranking with RL like in the OpenAI demo. I found that 8b models work well, but smaller models need multiple rewards that can be hacked, and don’t do so well. suyashss.github.io/llm/2025/03/08…
English
imho, GSM8k is not a proper RL demo task for these papers.
i propose the Sudoku task that @HrishbhDalal wrote a GRPO blogpost about.
it's not the same style of data that the Qwen team annealed on,
and it only works at ~7b+ scale (this applies to most complex RL tasks for LLMs)
机器之心 JIQIZHIXIN@jiqizhixin
GRPO just got a speed boost! Xiamen University introduced Completion Pruning Policy Optimization (CPPO), which significantly reduces the number of gradient calculations and updates. How fast? On GSM8K, it's 8.32× faster than GRPO, and on MATH, the speedup is 3.51×. 🚀🔥
English
More tagging typos: Thanks to @charles_irl
for the @modal_labs credits, and @HamelHusain and
@dan_s_becker for their LLM fine-tuning course.
English
Typo tagging @willccbb , thanks for the simple GRPO notebook.
English
🧬 Just published a new blog post on reinforcement fine-tuning (RFT) for causal gene identification using Llama models! We show how a fine-tuned 8B parameter model can match the performance of a 70B model on this task. 🧵 (1/8)
English
This implementation is laptop-friendly, so you can run it on your laptop in a couple of hours even if you don't have access to specialized GPUs.
English
Read more & try it out: 🔗 suyashss.github.io/vision-languag…
English
I built a tiny vision-language model using the MNIST digits dataset that can:
✅ Classify images → numbers
✅ Generate images from digit labels
It combines image tokenization & nanoGPT for training.
English
Suyash Shringarpure retweetledi
The genetic architecture of gene expression in individuals of African and European ancestry medrxiv.org/cgi/content/sh… #medRxiv
English
Suyash Shringarpure retweetledi
📌Publication alert: In a preprint from the @23andMeResearch team, we conducted one of the largest GWAS study of Long COVID and identified genetic links of chronic conditions with Long COVID.
medrxiv.org/content/10.110…
English