| # devops-300m-base — Recovered Base Checkpoint | |
| Recovered 2026-09-20 from /home/dkaiser/sft-training/devops-300m-prepped/ | |
| after the original outputs/devops-300m-4096-bf16-vast/ was lost in a | |
| working-copy rsync incident on big-ai. | |
| Provenance chain: | |
| 1. Trained on vast.ai (2x RTX 4090, bf16, 4000 steps, seq 4096) | |
| 2. Fetched to big-ai outputs/devops-300m-4096-bf16-vast/ | |
| 3. prep_for_sft.py converted to HF LlamaForCausalLM format (vocab 65539, | |
| +3 chat special tokens) -> sft-training/devops-300m-prepped/ | |
| 4. This copy = the prepped base weights (identical to trained checkpoint, | |
| HF-native format, loads via AutoModelForCausalLM) | |
| macpro mirror (/mnt/1tb_store/vastai/) was empty at recovery time. | |