All tags
Topic: "reward-models"
DeepSeek R1: o1-level open weights model and a simple recipe for upgrading 1.5B models to Sonnet/4o level
deepseek-r1 deepseek-v3 qwen-2.5 llama-3.1 llama-3.3-70b deepseek ollama qwen llama reinforcement-learning fine-tuning model-distillation model-optimization reasoning reward-models multi-response-sampling model-training
DeepSeek released DeepSeek R1, a significant upgrade over DeepSeek V3 from just three weeks prior, featuring 8 models including full-size 671B MoE models and multiple distillations from Qwen 2.5 and Llama 3.1/3.3. The models are MIT licensed, allowing finetuning and distillation. Pricing is notably cheaper than o1 by 27x-50x. The training process used GRPO (reward for correctness and style outcomes) without relying on PRM, MCTS, or reward models, focusing on reasoning improvements through reinforcement learning. Distilled models can run on Ollama and show strong capabilities like writing Manim code. The release emphasizes advances in reinforcement-learning, fine-tuning, and model-distillation with a novel RL framework from DeepSeekMath.
Did Nvidia's Nemotron 70B train on test?
nemotron-70b llama-3.1-70b llama-3.1 ministral-3b ministral-8b gpt-4o claude-3.5-sonnet claude-3.5 nvidia mistral-ai hugging-face zep benchmarking reinforcement-learning reward-models temporal-knowledge-graphs memory-layers context-windows model-releases open-source reach_vb philschmid swyx
NVIDIA's Nemotron-70B model has drawn scrutiny despite strong benchmark performances on Arena Hard, AlpacaEval, and MT-Bench, with some standard benchmarks like GPQA and MMLU Pro showing no improvement over the base Llama-3.1-70B. The new HelpSteer2-Preference dataset improves some benchmarks with minimal losses elsewhere. Meanwhile, Mistral released Ministral 3B and 8B models featuring 128k context length and outperforming Llama-3.1 and GPT-4o on various benchmarks under the Mistral Commercial License. NVIDIA's Nemotron 70B also surpasses GPT-4o and Claude-3.5-Sonnet on key benchmarks using RLHF (REINFORCE) training. Additionally, Zep introduced Graphiti, an open-source temporal knowledge graph memory layer for AI agents, built on Neo4j.
Nemotron-4-340B: NVIDIA's new large open models, built on syndata, great for syndata
nemotron-4-340b mixtral llama-3 gemini-1.5 gpt-4o mamba-2-hybrid-8b samba-3.8b-instruct dolphin-2.9.3 faro-yi-9b-dpo nvidia hugging-face mistral-ai llamaindex cohere gemini mistral synthetic-data model-alignment reward-models fine-tuning long-context model-scaling inference-speed mixture-of-agents open-source-models model-training instruction-following context-windows philipp-schmid bryan-catanzaro oleksii-kuchaiev rohanpaul_ai cognitivecompai _philschmid 01ai_yi
NVIDIA has scaled up its Nemotron-4 model from 15B to a massive 340B dense model, trained on 9T tokens, achieving performance comparable to GPT-4. The model alignment process uses over 98% synthetic data, with only about 20K human-annotated samples for fine-tuning and reward model training. The synthetic data generation pipeline is open-sourced, including synthetic prompts and preference data generation. The base and instruct versions outperform Mixtral and Llama 3, while the reward model ranks better than Gemini 1.5, Cohere, and GPT-4o. Other notable models include Mamba-2-Hybrid 8B, which is up to 8x faster than Transformers and excels on long-context tasks, Samba-3.8B-instruct for infinite context length with linear complexity, Dolphin-2.9.3 tiny models optimized for low-resource devices, and Faro Yi 9B DPO with a 200K context window running efficiently on 16GB VRAM. The Mixture-of-Agents technique boosts open-source LLMs beyond GPT-4 Omni on AlpacaEval 2.0.
Life after DPO (RewardBench)
gpt-3 gpt-4 gpt-5 gpt-6 llama-3-8b llama-3 claude-3 gemini x-ai openai mistral-ai anthropic cohere meta-ai-fair hugging-face nvidia reinforcement-learning-from-human-feedback direct-preference-optimization reward-models rewardbench language-model-history model-evaluation alignment-research preference-datasets personalization transformer-architecture nathan-lambert chris-manning elon-musk bindureddy rohanpaul_ai nearcyan
xAI raised $6 billion at a $24 billion valuation, positioning it among the most highly valued AI startups, with expectations to fund GPT-5 and GPT-6 class models. The RewardBench tool, developed by Nathan Lambert, evaluates reward models (RMs) for language models, showing Cohere's RMs outperforming open-source alternatives. The discussion highlights the evolution of language models from Claude Shannon's 1948 model to GPT-3 and beyond, emphasizing the role of RLHF (Reinforcement Learning from Human Feedback) and the newer DPO (Direct Preference Optimization) method. Notably, some Llama 3 8B reward model-focused models are currently outperforming GPT-4, Cohere, Gemini, and Claude on the RewardBench leaderboard, raising questions about reward hacking. Future alignment research directions include improving preference datasets, DPO techniques, and personalization in language models. The report also compares xAI's valuation with OpenAI, Mistral AI, and Anthropic, noting speculation about xAI's spending on Nvidia hardware.
1/16/2024: TIES-Merging
mixtral-8x7b nous-hermes-2 frankendpo-4x7b-bf16 thebloke hugging-face nous-research togethercompute oak-ridge-national-laboratory vast-ai runpod mixture-of-experts random-gate-routing quantization gptq exl2-quants reinforcement-learning-from-human-feedback supercomputing trillion-parameter-models ghost-attention model-fine-tuning reward-models sanjiwatsuki superking__ mrdragonfox _dampf kaltcit rombodawg technotech
TheBloke's Discord community actively discusses Mixture of Experts (MoE) models, focusing on random gate routing layers for training and the challenges of immediate model use. There is a robust debate on quantization methods, comparing GPTQ and EXL2 quants, with EXL2 noted for faster execution on specialized hardware. A new model, Nous Hermes 2, based on Mixtral 8x7B and trained with RLHF, claims benchmark superiority but shows some inconsistencies. The Frontier supercomputer at Oak Ridge National Laboratory is highlighted for training a trillion-parameter LLM with 14TB RAM, sparking discussions on open-sourcing government-funded AI research. Additionally, the application of ghost attention in the academicat model is explored, with mixed reactions from the community. "Random gate layer is good for training but not for immediate use," and "EXL2 might offer faster execution on specialized hardware," are key insights shared.