Hi @Teknium, I was thrilled to read your post's description. This sounds like the perfect role for me. Since you specifically asked for OSS contributions and GitHub links, I'll lead with those. I've poured my energy into Hugging Face's ecosystem, specifically post training optimizations and fixes that align spot on with your focus, plus other key projects. Here's the rundown: Hugging Face Transformers (my entry point. I tackled bugs, utils, and coverage boosts): - Fixed PaliGemma’s pad token unmasking during training: github.com/huggingface/trans… - Resolved audio classification bugs, including doc mismatches and runtime errors with top_k=None, plus test isolation fixes: github.com/huggingface/trans… - Made output_dir optional in TrainingArguments: github.com/huggingface/trans… - Added a reload utility for smoother dev workflows: github.com/huggingface/trans… Unsloth Optimization Challenges (I solved four tough ones and wrote them up on Medium): - Triton kernel for NF4 double dequantization to FP16/BF16 on T4 GPUs: medium.com/@indosambhav/unsl… - QLoRA FSDP2 integration for Llama 3.1 8B multi GPU fine tuning: medium.com/@indosambhav/unsl… - torch.compile without graph breaks for QLoRA in LLMs: medium.com/@indosambhav/unsl… - VRAM efficient backprop via chunked logits and generalized Cut Cross Entropy: medium.com/@indosambhav/unsl… Hugging Face PEFT (I've been contributing since October 2025, dropping impactful PRs on LoRA/DoRA/X LoRA enhancements. Fun fact: I'm the 3rd largest contributor there, right behind the maintainers, and #32 overall): - Negative weights support in LoRA merging, useful for task arithmetic and adapter subtraction: github.com/huggingface/peft/… - Embed_scale handling for DoRA, supporting scaled embeddings like Gemma3TextScaledWordEmbedding: github.com/huggingface/peft/… - Embed_scale handling for X LoRA: github.com/huggingface/peft/… - Scaled embeddings in TrainableTokensModel: github.com/huggingface/peft/… - Ensure_weight_tying for trainable_token_indices, which is upcoming for merge: github.com/huggingface/peft/… - 3 4x faster LoRA GA integration, also upcoming for merge: github.com/huggingface/peft/… Hugging Face Diffusers (as an MVP): - Submitted a fix for Qwen image encoding padding. The PR is up for review, so fingers crossed it merges soon: github.com/huggingface/diffu… - MVP program shoutout: github.com/huggingface/diffu… **Other Hugging Face Issues Raised** (red teaming OSS models, including OpenAI's on Kaggle): - Cross tokenizer distillation failures in TRL's GKD/MiniLLM trainers: github.com/huggingface/trl/i… - FA4 integration in Transformers: github.com/huggingface/trans… Developer Tools & More : - @pieces_for_developers: I'm the #3 contributor to pieces_cli_tool (github.com/pieces-app/cli-ag…) and #2 to Python OS SDK (github.com/pieces-app/pieces…). - Swarms, a multi agent orchestration library and LangChain alternative, where I'm an active contributor. - PECAn (Predictive Ecosystem Analyzer): I was a GSoC '24 mentee, working on downscaling ML outputs spatially and temporally, and I kept contributing post program. I really enjoy making AI models faster and more practical. Whether that's fine tuning LLMs at ByteLearn, a role I got thanks to a Kaggle contest, or sharing what I learn through open source. I'd love to chat when you have a moment. DM'ed you as well

Here is my version : I've poured my energy into Hugging Face's ecosystem, specifically post training optimizations and fixes that align spot on with your focus, plus other key projects. Here's the rundown: Hugging Face Transformers (my entry point. I tackled bugs, utils, and coverage boosts): - Fixed PaliGemma’s pad token unmasking during training: github.com/huggingface/trans… - Resolved audio classification bugs, including doc mismatches and runtime errors with top_k=None, plus test isolation fixes: github.com/huggingface/trans… - Made output_dir optional in TrainingArguments: github.com/huggingface/trans… - Added a reload utility for smoother dev workflows: github.com/huggingface/trans… Unsloth Optimization Challenges (I solved four tough ones and wrote them up on Medium): - Triton kernel for NF4 double dequantization to FP16/BF16 on T4 GPUs: medium.com/@indosambhav/unsl… - QLoRA FSDP2 integration for Llama 3.1 8B multi GPU fine tuning: medium.com/@indosambhav/unsl… - torch.compile without graph breaks for QLoRA in LLMs: medium.com/@indosambhav/unsl… - VRAM efficient backprop via chunked logits and generalized Cut Cross Entropy: medium.com/@indosambhav/unsl… Hugging Face PEFT (I've been contributing since October 2025, dropping impactful PRs on LoRA/DoRA/X LoRA enhancements. Fun fact: I'm the 3rd largest contributor there, right behind the maintainers, and #32 overall): - Negative weights support in LoRA merging, useful for task arithmetic and adapter subtraction: github.com/huggingface/peft/… - Embed_scale handling for DoRA, supporting scaled embeddings like Gemma3TextScaledWordEmbedding: github.com/huggingface/peft/… - Embed_scale handling for X LoRA: github.com/huggingface/peft/… - Scaled embeddings in TrainableTokensModel: github.com/huggingface/peft/… Ensure_weight_tying for trainable_token_indices, which is upcoming for merge: github.com/huggingface/peft/… - 3 4x faster LoRA GA integration, also upcoming for merge: github.com/huggingface/peft/… Hugging Face Diffusers (as an MVP): - Submitted a fix for Qwen image encoding padding. The PR is up for review, so fingers crossed it merges soon: github.com/huggingface/diffu… - MVP program shoutout: github.com/huggingface/diffu… Other Hugging Face Issues Raised (red teaming OSS models, including OpenAI's on Kaggle): - Cross tokenizer distillation failures in TRL's GKD/MiniLLM trainers: github.com/huggingface/trl/i… - FA4 integration in Transformers: github.com/huggingface/trans… Developer Tools & More : - @getpieces : I'm the #3 contributor to pieces_cli_tool (github.com/pieces-app/cli-ag…) and #2 to Python OS SDK (github.com/pieces-app/pieces…). - Swarms, a multi agent orchestration library and LangChain alternative, where I'm an active contributor. - PECAn (Predictive Ecosystem Analyzer): I was a GSoC '24 mentee, working on downscaling ML outputs spatially and temporally, and I kept contributing post program. I really enjoy making AI models faster and more practical. Whether that's fine tuning LLMs at ByteLearn, a role I got thanks to a Kaggle contest, or sharing what I learn through open source. I'd love to chat when you have a moment.

Hi @ClementDelangue , I was thrilled to read that you were hiring . Since you specifically asked for OSS contributions and GitHub links, I'll lead with those. I've poured my energy into Hugging Face's ecosystem, specifically post training optimizations and fixes that align spot on with your focus, plus other key projects. Here's the rundown: Hugging Face Transformers (my entry point. I tackled bugs, utils, and coverage boosts): - Fixed PaliGemma’s pad token unmasking during training: github.com/huggingface/trans… - Resolved audio classification bugs, including doc mismatches and runtime errors with top_k=None, plus test isolation fixes: github.com/huggingface/trans… - Made output_dir optional in TrainingArguments: github.com/huggingface/trans… - Added a reload utility for smoother dev workflows: github.com/huggingface/trans… Unsloth Optimization Challenges (I solved four tough ones and wrote them up on Medium): - Triton kernel for NF4 double dequantization to FP16/BF16 on T4 GPUs: medium.com/@indosambhav/unsl… - QLoRA FSDP2 integration for Llama 3.1 8B multi GPU fine tuning: medium.com/@indosambhav/unsl… - torch.compile without graph breaks for QLoRA in LLMs: medium.com/@indosambhav/unsl… - VRAM efficient backprop via chunked logits and generalized Cut Cross Entropy: medium.com/@indosambhav/unsl… Hugging Face PEFT (I've been contributing since October 2025, dropping impactful PRs on LoRA/DoRA/X LoRA enhancements. Fun fact: I'm the 3rd largest contributor there, right behind the maintainers, and #32 overall): - Negative weights support in LoRA merging, useful for task arithmetic and adapter subtraction: github.com/huggingface/peft/… - Embed_scale handling for DoRA, supporting scaled embeddings like Gemma3TextScaledWordEmbedding: github.com/huggingface/peft/… - Embed_scale handling for X LoRA: github.com/huggingface/peft/… - Scaled embeddings in TrainableTokensModel: github.com/huggingface/peft/… Ensure_weight_tying for trainable_token_indices, which is upcoming for merge: github.com/huggingface/peft/… - 3 4x faster LoRA GA integration, also upcoming for merge: github.com/huggingface/peft/… Hugging Face Diffusers (as an MVP): - Submitted a fix for Qwen image encoding padding. The PR is up for review, so fingers crossed it merges soon: github.com/huggingface/diffu… - MVP program shoutout: github.com/huggingface/diffu… Other Hugging Face Issues Raised (red teaming OSS models, including OpenAI's on Kaggle): - Cross tokenizer distillation failures in TRL's GKD/MiniLLM trainers: github.com/huggingface/trl/i… - FA4 integration in Transformers: github.com/huggingface/trans… Developer Tools & More : - @getpieces : I'm the #3 contributor to pieces_cli_tool (github.com/pieces-app/cli-ag…) and #2 to Python OS SDK (github.com/pieces-app/pieces…). - Swarms, a multi agent orchestration library and LangChain alternative, where I'm an active contributor. - PECAn (Predictive Ecosystem Analyzer): I was a GSoC '24 mentee, working on downscaling ML outputs spatially and temporally, and I kept contributing post program. I really enjoy making AI models faster and more practical. Whether that's fine tuning LLMs at ByteLearn, a role I got thanks to a Kaggle contest, or sharing what I learn through open source. I've DM'ed you as well and I'd love to chat when you have a moment.

Hi @QGallouedec, I was thrilled to read your post's description. This sounds like the perfect role for me. Since you specifically asked for OSS contributions and GitHub links, I'll lead with those. I've poured my energy into Hugging Face's ecosystem, specifically post training optimizations and fixes that align spot on with your focus, plus other key projects. Here's the rundown: Hugging Face Transformers (my entry point. I tackled bugs, utils, and coverage boosts): - Fixed PaliGemma’s pad token unmasking during training: github.com/huggingface/trans… - Resolved audio classification bugs, including doc mismatches and runtime errors with top_k=None, plus test isolation fixes: github.com/huggingface/trans… - Made output_dir optional in TrainingArguments: github.com/huggingface/trans… - Added a reload utility for smoother dev workflows: github.com/huggingface/trans… Unsloth Optimization Challenges (I solved four tough ones and wrote them up on Medium): - Triton kernel for NF4 double dequantization to FP16/BF16 on T4 GPUs: medium.com/@indosambhav/unsl… - QLoRA FSDP2 integration for Llama 3.1 8B multi GPU fine tuning: medium.com/@indosambhav/unsl… - torch.compile without graph breaks for QLoRA in LLMs: medium.com/@indosambhav/unsl… - VRAM efficient backprop via chunked logits and generalized Cut Cross Entropy: medium.com/@indosambhav/unsl… Hugging Face PEFT (I've been contributing since October 2025, dropping impactful PRs on LoRA/DoRA/X LoRA enhancements. Fun fact: I'm the 3rd largest contributor there, right behind the maintainers, and #32 overall): - Negative weights support in LoRA merging, useful for task arithmetic and adapter subtraction: github.com/huggingface/peft/… - Embed_scale handling for DoRA, supporting scaled embeddings like Gemma3TextScaledWordEmbedding: github.com/huggingface/peft/… - Embed_scale handling for X LoRA: github.com/huggingface/peft/… - Scaled embeddings in TrainableTokensModel: github.com/huggingface/peft/… - Ensure_weight_tying for trainable_token_indices, which is upcoming for merge: github.com/huggingface/peft/… - 3 4x faster LoRA GA integration, also upcoming for merge: github.com/huggingface/peft/… Hugging Face Diffusers (as an MVP): - Submitted a fix for Qwen image encoding padding. The PR is up for review, so fingers crossed it merges soon: github.com/huggingface/diffu… - MVP program shoutout: github.com/huggingface/diffu… **Other Hugging Face Issues Raised** (red teaming OSS models, including OpenAI's on Kaggle): - Cross tokenizer distillation failures in TRL's GKD/MiniLLM trainers: github.com/huggingface/trl/i… - FA4 integration in Transformers: github.com/huggingface/trans… Developer Tools & More : - @pieces_for_developers: I'm the #3 contributor to pieces_cli_tool (github.com/pieces-app/cli-ag…) and #2 to Python OS SDK (github.com/pieces-app/pieces…). - Swarms, a multi agent orchestration library and LangChain alternative, where I'm an active contributor. - PECAn (Predictive Ecosystem Analyzer): I was a GSoC '24 mentee, working on downscaling ML outputs spatially and temporally, and I kept contributing post program. I really enjoy making AI models faster and more practical. Whether that's fine tuning LLMs at ByteLearn, a role I got thanks to a Kaggle contest, or sharing what I learn through open source. I'd love to chat when you have a moment.

There's an ongoing discourse on whether #AI is a boon or a bane in educational spaces. And while it'll certainly continue to be a point of contention, US-based Bytelearn is offering a niche learning advantage in this space--read all about it here: entrepreneur.com/en-ae/entre…

Thrilled to receive "ET Young Leaders Award" #ETYLA & #ByteLearn being recognized as an #EmergingStartup! 🎉💫 Grateful to my family, wife, cofounder @AdityaSinghal, team & mentors for their unwavering support! 🙏 #EdTech #Leadership #Innovation

15 AI tools to help you learn efficiently: 1. Poised 2. TutorAI 3. Eightify 4. Vocalo 5. Socratic 6. ChatPDF 7. Fluently 8. Kippy AI 9. YouLearn 10. Langotalk 11. Mindgrasp 12. ByteLearn 13. elsaspeak 14. MathGPTPro 15. Learniverse AI What tool did I miss?

Bytelearn is hiring for Machine Learning Engineer Intern Stipend: ₹ 25K - 27K per month Location: Remote Apply here: cuvette.tech/app/public/inte… #jobs #hiring

Applied for Front-end Intern role at bytelearn. Got shortlisted. Got the assignment. Submitted it. Waiting for their response. 👍

Remote Frontend Internship #hiring 🚨 Bytelearn is hiring Frontend Developers🚩 Stipend ✨: Rs 25K / month Eligible: 2026, 2025, 2024, 2023, 2022 Skills: HTML, CSS, Javascript Like & comment "Interview" for better reach 🚀 Link: cuvette.tech/app/public/inte… #tech #frontend

Remote Internship Alert! 🚨 Bytelearn Pvt Ltd is hiring “Frontend Developer Interns” Link: bit.ly/remote-intern-ship Stipend: Rs. 20K - 25K / month Eligible: 2026, 2025, 2024, 2023, 2022 Skills: HTML, CSS, Javascript #Frontend #internship #opportunity

Remote Internship Alert! 🚨 Bytelearn Pvt Ltd is hiring “Frontend Developer Interns” Stipend: Rs. 20K - 25K / month Eligible: 2026, 2025, 2024, 2023, 2022 Skills: HTML, CSS, Javascript Comment “Interested” & I’ll DM you the link. #tech #frontend #Web3 #Engineering

Bytelearn is hiring for multiple intern roles Stipend: 15k-30k per month Location: Remote Apply here for Machine Learning Engineer Intern: cuvette.tech/app/public/inte… Apply here for Frontend Developer Intern: cuvette.tech/app/public/inte… #intern #hiring

2. Machine Learning Engineer Internship Bytelearn | Work From Home Batch : 2023 Stipend Per Month : ₹ 25K - 30K Apply:lnkd.in/dyFsdfnF

Working on One-Step Equations? Try a free ByteLearn lesson plan on One-Step Equations. Includes worksheet, quizzes, practice problems and more. hubs.la/Q01H_k0B0 #maths #middleschool #6thgrade

At #ByteLearn we understand that not all students learn at the same pace. That's why we offer scaffolded #math lessons to help students build a strong foundation. #maths

At ByteLearn, we're committed to helping every student achieve their full potential in #math. Join us today and discover how our scaffolded approach can transform your child's math skills! #maths

Math doesn't have to be scary! #ByteLearn is dedicated to making math accessible and enjoyable for all middle school students. #maths

Great to hear that Bytelearn is making an impact at #ATOMIC2023 Math conference! It's exciting to see companies like yours innovating and changing the way students learn and understand Math. Keep up the good work! #EdTech #MathEducation

Bytelearn is at #ATOMIC2023 Math conference. We are changing the way students learn and understand #Math