Is SFT Dead? How Meta only uses 13 Parameters to Learning to Reason скачать в хорошем качестве

Is SFT Dead? How Meta only uses 13 Parameters to Learning to Reason

What if I told you that a 7B model like Qwen2.5-7B could jump from 76% to 91% accuracy… using just 26 bytes of trainable data? That’s smaller than a tweet. In this video, we break down Meta’s groundbreaking paper, Meta’s “Learning to Reason in 13 Parameters.” We explore how TinyLoRA challenges everything we thought we knew about fine-tuning large language models. We’ll cover: Why 13 parameters can outperform traditional LoRA setups Why Reinforcement Learning (GRPO) crushes Supervised Fine-Tuning for reasoning The shocking “Inverse Scaling Law” that suggests bigger models may need less training If you care about alignment, local LLMs, or the future of AI efficiency, this one will rewire how you think about model steering. Join my AI newsletter: https://upaspro.com/newsletter/ More information: https://upaspro.com/is-sft-dead-how-m... 👇 Timestamps: 00:00- Train Qwen 2.5 with 26 bytes 02:02- 1- Myth of Capacity 04:13- 2- Signal-to-Noise Ratio 06:28- 3- Inverse Scaling Law 07:09- Recap #AI #LLM #TinyLoRA #ReinforcementLearning #MetaAI #MachineLearning #OpenSourceAI #Alignment #Qwen #DeepLearning