lilywchen/lucky-initialization-atlas-100m-v2-checkpoints

Lucky initialization atlas v2 checkpoints

Private, continuously synchronized model-only checkpoints for the prospectively frozen OLMo-3-ladder 100M initialization atlas. Every visible checkpoint was converted with the pinned olmo_core.nn.hf.save_hf_model, serialized with its pinned tokenizer, and SHA-256 manifested before upload.

Load a specific model with standard Transformers APIs:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "lilywchen/lucky-initialization-atlas-100m-v2-checkpoints"
subfolder = "models/<run-id>/step-006210"
model = AutoModelForCausalLM.from_pretrained(repo, subfolder=subfolder)
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=subfolder)

These are model-only checkpoints. Exact in-progress recovery uses separate local full-state checkpoints; optimizer and RNG states are intentionally not published. The study's primary claims are conditional on its frozen FineWeb corpus/order, OLMo-core recipe, evaluation sets, and B200 runtime.