All tags
Company: "dbrx"
Talaria: Apple's new MLOps Superweapon
gemma mixtral phi dbrx apple google mistral-ai microsoft mosaic quantization on-device-ai adapter-models model-optimization model-latency lossless-quantization low-bit-palletization token-generation model-benchmarking human-evaluation craig-federighi andrej-karpathy
Apple Intelligence introduces a small (~3B parameters) on-device model and a larger server model running on Apple Silicon with Private Cloud Compute, aiming to surpass Google Gemma, Mistral Mixtral, Microsoft Phi, and Mosaic DBRX. The on-device model features a novel lossless quantization strategy using mixed 2-bit and 4-bit LoRA adapters averaging 3.5 bits-per-weight, enabling dynamic adapter hot-swapping and efficient memory management. Apple credits the Talaria tool for optimizing quantization and model latency, achieving about 0.6 ms time-to-first-token latency and 30 tokens per second generation rate on iPhone 15 Pro. Apple focuses on an "adapter for everything" strategy with initial deployment on SiriKit and App Intents. Performance benchmarks rely on human graders, emphasizing consumer-level adequacy over academic dominance. The Apple ML blog also mentions an Xcode code-focused model and a diffusion model for Genmoji.
FineWeb: 15T Tokens, 12 years of CommonCrawl (deduped and filtered, you're welcome)
llama-3-70b llama-3 wizardlm-2-8x22b claude-opus mistral-8x7b gpt-4 huggingface meta-ai-fair dbrx reka-ai mistral-ai lmsys openai datasets benchmarking quantization zero-shot-learning reasoning code-error-detection token-generation security
2024 has seen a significant increase in dataset sizes for training large language models, with Redpajama 2 offering up to 30T tokens, DBRX at 12T tokens, Reka Core/Flash/Edge with 5T tokens, and Llama 3 trained on 15T tokens. Huggingface released an open dataset containing 15T tokens from 12 years of filtered CommonCrawl data, enabling training of models like Llama 3 if compute resources are available. On Reddit, WizardLM-2-8x22b outperformed other open LLMs including Llama-3-70b-instruct in reasoning and math benchmarks. Claude Opus demonstrated strong zero-shot code error spotting, surpassing Llama 3. Benchmarks revealed limitations in the LMSYS chatbot leaderboard due to instruction-tuned models gaming the system, and a new RAG benchmark showed Llama 3 70B underperforming compared to GPT-4, while Mistral 8x7B remained strong. Efficient quantized versions of Llama 3 models are available on Huggingface, with users reporting token generation limits around 9600 tokens on a 3090 GPU. Safety concerns include a UK sex offender banned from AI tool usage and GPT-4 demonstrating an 87% success rate exploiting real vulnerabilities, raising security concerns.
Jamba: Mixture of Architectures dethrones Mixtral
jamba dbrx mixtral animatediff fastsd sdxs512-0.9 b-lora supir ai21-labs databricks together-ai hugging-face midjourney mixture-of-experts model-architecture context-windows model-optimization fine-tuning image-generation video-generation cpu-optimization style-content-separation high-resolution-upscaling
AI21 labs released Jamba, a 52B parameter MoE model with 256K context length and open weights under Apache 2.0 license, optimized for single A100 GPU performance. It features a unique blocks-and-layers architecture combining transformer and MoE layers, competing with models like Mixtral. Meanwhile, Databricks introduced DBRX, a 36B active parameter MoE model trained on 12T tokens, noted as a new standard for open LLMs. In image generation, advancements include Animatediff for video-quality image generation and FastSD CPU v1.0.0 beta 28 enabling ultra-fast image generation on CPUs. Other innovations involve style-content separation using B-LoRA and improvements in high-resolution image upscaling with SUPIR.
DBRX: Best open model (just not most efficient)
dbrx grok mixtral llama-2 mpt-7b gpt-4 databricks hugging-face mistral-ai mosaicml openai mixture-of-experts model-efficiency tokenization model-training code-generation model-architecture open-source-models benchmarking fine-tuning
Databricks Mosaic has released a new open-source model called DBRX that outperforms Grok, Mixtral, and Llama2 on evaluations while being about 2x more efficient than Llama2 and Grok. The model was trained on 12 trillion tokens using 3,000 H100 GPUs over 2 months, with an estimated compute cost of $10 million. It uses OpenAI's 100k tiktoken tokenizer and shows strong zero-shot code generation performance, even beating GPT-4 on the Humaneval benchmark. DBRX also upstreamed work to MegaBlocks open source. Despite its scale and efficiency, DBRX's performance on MMLU is only slightly better than Mixtral, raising questions about its scaling efficiency. The focus of DBRX is on enabling users to train models efficiently, with MoE training being about 2x more FLOP-efficient than dense models, achieving similar quality with nearly 4x less compute than previous MPT models. This release is part of the ongoing competition for open-source AI leadership, including models like Dolly, MPT, and Mistral. "If it activates 36B params, the model's perf should be equivalent to a 72B dense model or even 80B," says Qwen's tech lead.