Aflah 🍉🕊️
@aflah02101.bsky.social
Research Software Engineer @MPI-SWS • OSS @EleutherAI• Prev @Goldman Sachs, @LCS2, GSoC @TensorFlow • IIIT Delhi '24 • #CEASEFIRENOW 🕊️
Reposted by Aflah 🍉🕊️
Announcing 🔭Hubble, a suite of open-source LLMs to advance the study of memorization!
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
October 24, 2025 at 6:21 PM
Announcing 🔭Hubble, a suite of open-source LLMs to advance the study of memorization!
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
Something we've been cooking for the past year
Looking forward to what the community builds and studies using these models!!
Looking forward to what the community builds and studies using these models!!
Announcing 🔭Hubble, a suite of open-source LLMs to advance the study of memorization!
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
Pretrained 1B/8B param models, with controlled insertion of texts designed to emulate key memorization risks: copyright (e.g., book passages), privacy (e.g., synthetic biographies), and test set contamination
October 24, 2025 at 6:23 PM
Something we've been cooking for the past year
Looking forward to what the community builds and studies using these models!!
Looking forward to what the community builds and studies using these models!!
So academics in the US can raise their voice but only when their funding is under fire. Please don't pretend to have more freedom when you're too scared to use it
June 14, 2025 at 9:08 PM
So academics in the US can raise their voice but only when their funding is under fire. Please don't pretend to have more freedom when you're too scared to use it
I was just looking at the stats for the resources github I used to maintain during my undergrad for the different courses and its so nice to see still so many regular visitors and contributors
February 21, 2025 at 1:50 PM
I was just looking at the stats for the resources github I used to maintain during my undergrad for the different courses and its so nice to see still so many regular visitors and contributors
The best skill I've learnt from my parents is never hesitating to spend on books. I see people see stuff that will give them high ROI (knowing well it will themselves) and not buy it saying it's too expensive when they spend like 5x the amount on drinks
February 16, 2025 at 8:28 PM
The best skill I've learnt from my parents is never hesitating to spend on books. I see people see stuff that will give them high ROI (knowing well it will themselves) and not buy it saying it's too expensive when they spend like 5x the amount on drinks
It is so funny when I see some seniors/batchmates give tips for cracking FAANG companies on social media. I know who gave the test on your behalf, please keep your BS to yourself 😑
February 16, 2025 at 7:44 AM
It is so funny when I see some seniors/batchmates give tips for cracking FAANG companies on social media. I know who gave the test on your behalf, please keep your BS to yourself 😑
My blogpost on running multinode slurm inference with SGLang is now adapted into the docs!
Docs - github.com/sgl-project/...
Blog - aflah02.substack.com/p/multi-node...
Docs - github.com/sgl-project/...
Blog - aflah02.substack.com/p/multi-node...
github.com
February 14, 2025 at 10:34 AM
My blogpost on running multinode slurm inference with SGLang is now adapted into the docs!
Docs - github.com/sgl-project/...
Blog - aflah02.substack.com/p/multi-node...
Docs - github.com/sgl-project/...
Blog - aflah02.substack.com/p/multi-node...
It is so funny that the same people this podcaster gave platform, helped them peddle pseudoscience and their hateful narrative without push back are now the ones leading the witch-hunt against him.
"... Then they came for me—and there was no one left to speak for me."
"... Then they came for me—and there was no one left to speak for me."
February 11, 2025 at 6:47 AM
It is so funny that the same people this podcaster gave platform, helped them peddle pseudoscience and their hateful narrative without push back are now the ones leading the witch-hunt against him.
"... Then they came for me—and there was no one left to speak for me."
"... Then they came for me—and there was no one left to speak for me."
I really like that the cop pulled the plug on Ed's street performance. Why should people who are just trying to go about their work in one of the most traffic prone cities have to pay the price of such impromptu performances? I could care less for any singer and a lot of people feel the same
February 10, 2025 at 6:40 AM
I really like that the cop pulled the plug on Ed's street performance. Why should people who are just trying to go about their work in one of the most traffic prone cities have to pay the price of such impromptu performances? I could care less for any singer and a lot of people feel the same
Book Fair 2025 Haul
February 9, 2025 at 3:16 PM
Book Fair 2025 Haul
Once I get done with this pretraining project I'm defo going to make a video series on how to pretrain LMs using NeoX with all the things I wish I knew before I started. Pretraining isn't hard or unreachable especially for smaller models with Academia scale compute!!
February 8, 2025 at 10:04 AM
Once I get done with this pretraining project I'm defo going to make a video series on how to pretrain LMs using NeoX with all the things I wish I knew before I started. Pretraining isn't hard or unreachable especially for smaller models with Academia scale compute!!
I hate Anthropic's approach to AI safety
February 7, 2025 at 8:49 PM
I hate Anthropic's approach to AI safety
When RL finally clicks for you it's such a great moment
February 6, 2025 at 7:14 PM
When RL finally clicks for you it's such a great moment
If you want to understand the current SOTA post training regime just read Tulu 1, 2, 2.5 and 3 + DeepSeek Math, V3 and R1. Both these series gel so seamlessly across their papers.
February 3, 2025 at 7:05 PM
If you want to understand the current SOTA post training regime just read Tulu 1, 2, 2.5 and 3 + DeepSeek Math, V3 and R1. Both these series gel so seamlessly across their papers.
Spent the entire day hard optimising pretraining throughput
Feeling like I did some real work after a while when all I did was turn flags on and off and look at wandb logs 🤓
Feeling like I did some real work after a while when all I did was turn flags on and off and look at wandb logs 🤓
February 2, 2025 at 8:56 PM
Spent the entire day hard optimising pretraining throughput
Feeling like I did some real work after a while when all I did was turn flags on and off and look at wandb logs 🤓
Feeling like I did some real work after a while when all I did was turn flags on and off and look at wandb logs 🤓
Recently saw an interview of a senior researcher from DeepSeek and realised that my first research project involved using her codebase as a baseline lol
P.s. I could never get the code to run lol as setting up an env for TF1 was borderline impossible without proper pinning
P.s. I could never get the code to run lol as setting up an env for TF1 was borderline impossible without proper pinning
February 1, 2025 at 9:43 PM
Recently saw an interview of a senior researcher from DeepSeek and realised that my first research project involved using her codebase as a baseline lol
P.s. I could never get the code to run lol as setting up an env for TF1 was borderline impossible without proper pinning
P.s. I could never get the code to run lol as setting up an env for TF1 was borderline impossible without proper pinning
Is something wrong with bluesky as I keep getting the same notification for someone following me back or is the person just following and unfollowing me 😂
December 7, 2024 at 10:12 PM
Is something wrong with bluesky as I keep getting the same notification for someone following me back or is the person just following and unfollowing me 😂
Reposted by Aflah 🍉🕊️
I'm excited to kick off my Bluesky presence with wonderful news: Our paper "Reference-Based Metrics Are Biased Against Blind and Low-Vision Users' Image Description Preferences" won a Best Paper Award at the NLP for Positive Impact Workshop at EMNLP! Read it here: aclanthology.org/2024.nlp4pi-...
Reference-Based Metrics Are Biased Against Blind and Low-Vision Users’ Image Description Preferences
Rhea Kapur, Elisa Kreiss. Proceedings of the Third Workshop on NLP for Positive Impact. 2024.
aclanthology.org
November 24, 2024 at 6:39 PM
I'm excited to kick off my Bluesky presence with wonderful news: Our paper "Reference-Based Metrics Are Biased Against Blind and Low-Vision Users' Image Description Preferences" won a Best Paper Award at the NLP for Positive Impact Workshop at EMNLP! Read it here: aclanthology.org/2024.nlp4pi-...