All tags
Topic: "knowledge-distillation"
Reasoning Models are Near-Superhuman Coders (OpenAI IOI, Nvidia Kernels)
o3 o1 o3-mini deepseek-r1 qwen-2.5 openthinker openai nvidia ollama elevenlabs sakana-ai apple reinforcement-learning gpu-kernel-optimization fine-tuning knowledge-distillation scaling-laws chain-of-thought-reasoning model-accessibility alex-wei karpathy abacaj awnihannun
o3 model achieved a gold medal at the 2024 IOI and ranks in the 99.8 percentile on Codeforces, outperforming most humans with reinforcement learning (RL) methods proving superior to inductive bias approaches. Nvidia's DeepSeek-R1 autonomously generates GPU kernels that surpass some expert-engineered kernels, showcasing simple yet effective AI-driven optimization. OpenAI updated o1 and o3-mini models to support file and image uploads in ChatGPT and released DeepResearch, a powerful research assistant based on the o3 model with RL for deep chain-of-thought reasoning. Ollama introduced OpenThinker models fine-tuned from Qwen2.5, outperforming some DeepSeek-R1 distillation models. ElevenLabs grew into a $3.3 billion company specializing in AI voice synthesis without open-sourcing their technology. Research highlights include Sakana AI Labs' TAID knowledge distillation method receiving a Spotlight at ICLR 2025, and Apple's work on scaling laws for mixture-of-experts (MoEs). The importance of open-source AI for scientific discovery was also emphasized.
Nvidia Minitron: LLM Pruning and Distillation updated for Llama 3.1
llama-3-1-8b llama-3-1 jamba-1.5 claude-3 dracarys-70b dracarys-72b mistral-nemo-minitron-8b mistral-7b nvidia meta-ai-fair ai21-labs anthropic hugging-face pruning knowledge-distillation weight-pruning activation-based-pruning width-pruning kl-divergence teacher-correction prompt-optimization multilinguality long-context mixture-of-experts model-fine-tuning
Nvidia and Meta researchers updated their Llama 3 results with a paper demonstrating the effectiveness of combining weight pruning and knowledge distillation to reduce training costs by training only the largest model from scratch and deriving smaller models via pruning and distillation. The process involves teacher correction, activation-based pruning (favoring width pruning), and retraining with distillation using KL Divergence loss, resulting in better-performing models at comparable sizes. However, distillation incurs some accuracy tradeoffs. Additionally, AI21 Labs launched Jamba 1.5, a hybrid SSM-Transformer MoE model with large context windows and multilingual support. Anthropic updated Claude 3 with LaTeX rendering and prompt caching. An open-source coding-focused LLM, Dracarys, was released in 70B and 72B sizes, showing improved coding performance. The Mistral Nemo Minitron 8B model outperforms Llama 3.1 8B and Mistral 7B on the Hugging Face leaderboard, highlighting pruning and distillation benefits. Research on prompt optimization reveals the complexity of prompt search spaces and the surprising effectiveness of simple algorithms like AutoPrompt/GCG.
Gemma 2 2B + Scope + Shield
gemma-2b gemma-2-9b gemma-2-27b llama-3-1-405b sam-2 gpt-3.5 vicuna alpacaeval g-eval google-deepmind anthropic meta-ai-fair openai perplexity-ai nvidia lmsys knowledge-distillation leaderboards model-interpretability finetuning harm-detection video-segmentation voice publishers-program robotics-data-scaling quantization llm-evaluation prompt-engineering
Gemma 2B, a 2 billion parameter model trained on 2 trillion tokens and distilled from a larger unnamed LLM, has been released by Google DeepMind and shows strong leaderboard performance despite weaknesses in math. The Gemma series, including 9B and 27B models, has gained popularity since its June release. The team also released 400 SAEs for interpretability, inspired by Anthropic's research. A finetuned classifier called ShieldGemma outperforms Meta's LlamaGuard in harm detection. Meanwhile, Meta AI announced Llama-3.1-405B reaching #3 on the Overall Arena leaderboard, and released SAM 2, a video and image segmentation model with significant speed improvements. OpenAI is rolling out an advanced Voice Mode to Plus users. Perplexity AI launched a Publishers Program with major media partners and a status page. NVIDIA introduced Project GR00T for scaling robot data using Apple Vision Pro and generative simulation. Interest in quantization for compressing LLMs is growing, and LLM-as-a-Judge implementations from Vicuna, AlpacaEval, and G-Eval highlight the effectiveness of simple prompts and domain-specific evaluation.
Gemma 2: The Open Model for Everyone
gemma-2 qwen-72b mixtral-8x22b-instruct claude-3.5-sonnet google-deepmind alibaba mistral-ai anthropic knowledge-distillation attention-mechanisms multilingual-models multimodality model-training model-optimization memory-optimization fine-tuning kathleen-kenealy daniel-han
Gemma 2, a 27B parameter model from google-deepmind, was released with innovations like 1:1 local-global attention alternation and logit soft-capping, leveraging knowledge distillation to train smaller models on over 50× the compute-optimal token quantity. The model supports multilingual and multimodal capabilities, with fine-tuning success on over 200 Indic language variants. The Open LLM Leaderboard highlights alibaba's Qwen 72B as the top model, with mistral-ai's Mixtral-8x22B-Instruct also ranking highly. Anthropic launched Claude 3.5 Sonnet, improving intelligence at mid-tier cost and speed. Research on eliminating matrix multiplication in LLMs promises significant memory savings without performance loss. Kathleen Kenealy and Daniel Han provided insights on Gemma 2's tokenizer and attention scaling respectively.
GPT-4o: the new SOTA-EVERYTHING Frontier model (GPT4T version)
gpt-4o gpt-3.5 llama-3 openai hugging-face nous-research eleutherai hazyresearch real-time-reasoning coding-capabilities fine-tuning knowledge-distillation hardware-optimization quantization multimodality mixture-of-experts efficient-attention model-scaling depth-upscaling transformer-architecture gpu-optimization prompt-engineering
OpenAI launched GPT-4o, a frontier model supporting real-time reasoning across audio, vision, and text, now free for all ChatGPT users with enhanced coding capabilities and upcoming advanced voice and video features. Discussions cover open-source LLMs like Llama 3, fine-tuning techniques including knowledge distillation for GPT-3.5, and hardware optimization strategies such as quantization. Emerging architectures include multimodal integrations with ChatGPT voice and Open Interpreter API, Mixture of Experts models combining autoregressive and diffusion approaches, and novel designs like the YOCO architecture and ThunderKittens DSL for efficient GPU use. Research advances in efficient attention methods like Conv-Basis using FFT and model scaling techniques such as depth upscaling were also highlighted.