YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
control-pretrain-30b-baseline — raw Megatron checkpoint backup
Off-cluster backup of the GEOD-201 control-pretraining 30B baseline campaign's raw
torch_dist checkpoints (optimizer state included), synced from Isambard:
/projects/a5k/public/data_cwtice.a5k/checkpoints/megatron/control_pretraining/
Layout: <stage>/iter_XXXXXXX/... for stages pretrain / midtrain / sft.
These are NOT HuggingFace-format models — they are Megatron distributed checkpoints
(Nemotron 3 Nano 30B-A3B architecture, trained from scratch) kept for disaster
recovery and resume. HF-format exports of selected iterations are published
separately via the conversion pipeline.
This repo is public and will grow to ~5-6 TB over the campaign.
Sync watcher: data_cwtice.a5k/scripts/hf_ckpt_sync.py on Isambard.
(Moved here from camgeodesic/ 2026-08-23 once org billing was enabled.)