Revision	Training step	Notes
`step-000040`	40	SFT checkpoint
`step-000080`	80	SFT checkpoint
`step-000160`	160	early / donor
`step-000240`	240	SFT checkpoint
`step-000320`	320	SFT checkpoint
`step-000400`	400	SFT checkpoint
`step-000480`	480	SFT checkpoint
`step-000640`	640	SFT checkpoint
`step-000720`	720	SFT checkpoint
`step-000760`	760	near-peak
`step-000800`	800	near-peak

Load

python

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "PHIMemo/llama31-8b-instruct-sft-balanced-3k"
rev = "step-000800"  # or any revision above
tok = AutoTokenizer.from_pretrained(repo, revision=rev)
model = AutoModelForCausalLM.from_pretrained(repo, revision=rev, torch_dtype="auto", device_map="auto")

Notes

Synthetic dossiers only (not real PHI).
Weights only (model.safetensors + configs); optimizer / trainer state not uploaded.
Gated base model terms still apply for Llama 3.1.