Marcolini/nanochat-d24-base-champion

docs: cite the eval log and the skipped-example count

docs: explain the meta.json val_bpb gap

docs: link the published r24 checkpoint

docs: name the real bpb eval path (split-tokens, not skip5120)

Link the companion SFT data mixes dataset

Point clone URL at nanochat-on-rtx4070

Correct usage section: no bundled tokenizer, real load_example.py invocation, stored dtype

Rewrite model card: consumer-GPU provenance, eval table, honest limitations

Upload folder using huggingface_hub

initial commit