Haha cool to see my graph in a German newspaper (Frankfurter Allgemeine Zeitung) 🙂 (But R.I.P. Falcon 180B)
@maximelabonne
-
Large AI Model Shows Redundancy in Weight Parameters
By
–
It looks pretty good but expected even better from such a big model (so much redundancy in these weights)
-
Jeremy Howard’s Quest to Upgrade Software Engineering
By
–
Please stop this man, @jeremyphoward is on a personal quest to upgrade every branch of software engineering
-
BOS Token Configuration Issue in Language Models
By
–
It's a good question. You're right, they're added in the background. In this particular example, I noticed that im_start wasn't properly added however. We still use "" as "bos_token".
-
Victor Sungo’s Technique Now Widely Adopted in AI
By
–
Congrats @victorsungo
! It must feel nice to see your technique used everywhere now 🙂 -
Apple Intelligence: 3B On-Device and 70B Server Models
By
–
Post-training in the Apple Intelligence paper – Two models: – On-device: ~3B param with task-specific LoRA adapters – Server: ~70B (my estimation), almost no details – Human-annotated and synthetic data: – Mathematics: @WizardLM_AI
-like evol strategy to create -
Distilabel for Synthetic Data Generation: Real Use Case
By
–
I actually talk about a "real" use case in my book. For synthetic data generation, I recommend checking distilabel by @argilla_io
-
Unsloth TRL Rewrite: Improved Performance for AI Model Training
By
–
Thanks Daniel! I originally had the entire code in classic TRL and decided to rewrite it with Unsloth two days ago. I don't regret it, it's super performant and convenient. Great work! 🙂
-
Cute Anime Llama Warrior Illustration Generated by AI
By
–
"Draw the illustration of a cute young llama with a sword and shield sleeping under a tree. The llama has arms and legs but is not human. Anime movie style, Japanese author, sakuga. Landscape format."
-
Fine-tuning LLMs with TRL and Axolotl alternatives
By
–
In Unsloth? I don't know, I'm not part of the dev team. You can still fine-tune with TRL or Axolotl.