--- base_model: Qwen/Qwen2.5-0.5B library_name: transformers tags: - qwen2 - pretraining - mu-greats --- # muon-gradientMatching Qwen2.5-0.5B models pretrained from scratch on `openwebtext-100k`, from the Mu-GREATS subset-selection / Muon gradient-matching experiments. Each run is a subfolder; load one with: ```python from transformers import AutoModelForCausalLM, AutoTokenizer repo = "efficiency-learning/muon-gradientMatching" sub = "full_muon_lr0.001" # any folder below model = AutoModelForCausalLM.from_pretrained(repo, subfolder=sub) tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=sub) ``` | Subfolder | Optimizer | LR | Mode | Step | dtype | |---|---|---|---|---|---| | `full_adamw_lr0.0001` | adamw | 0.0001 | full | 6317 | ? | | `full_adamw_lr0.0005` | adamw | 0.0005 | full | 6317 | ? | | `full_adamw_lr0.001` | adamw | 0.001 | full | 6317 | ? | | `full_adamw_lr0.005` | adamw | 0.005 | full | 6317 | ? | | `full_adamw_lr5e-05` | adamw | 5e-05 | full | 6317 | ? | | `full_muon_lr0.0001` | muon | 0.0001 | full | 6317 | ? | | `full_muon_lr0.0002` | muon | 0.0002 | full | 6317 | ? | | `full_muon_lr0.0005` | muon | 0.0005 | full | 6317 | ? | | `full_muon_lr0.001` | muon | 0.001 | full | 6317 | ? | | `full_muon_lr0.005` | muon | 0.005 | full | 6317 | ? | | `full_muon_lr5e-05` | muon | 5e-05 | full | 6317 | ? | | `random_muon_lr0.001` | muon | 0.001 | random | 6317 | ? |