OVD checkpoints

Paper checkpoints grouped by training-data size, method, and step.

• 1-data/{rlvr,full,hes,random}/stepN/: one fixed training problem.
• 128-data/{rlvr,full,hes,random}/stepN/: 128 distinct training problems.
• web/: Web experiment checkpoints, pending identity verification.

Each uploaded checkpoint has its own config, tokenizer, weights, and provenance. Upload coverage is recorded in upload_manifest.json; not all planned folders are available yet. This first batch covers 12 OVD checkpoints under 128-data and eight RLVR checkpoints across both data settings. These are evaluated historical models, not outputs of the recently repaired code. Optimizer state is not included.

Load a completed checkpoint with AutoModelForCausalLM.from_pretrained("menik1126/OVD", subfolder="128-data/rlvr/step300"). Use the same subfolder when loading the tokenizer.