config.json

the model repo, simple: the 1.1 container, the master weights, a short card; the technical read is on GitHub

model card: library_name shadow, so .shdw downloads are counted

config.json: the model's shape, so downloads are counted

harnesses: video memory, records desk, draft model, memory server; code, write-ups and recordings, no weights

v1.1.1 container: the archive blobs back in, extended to the new rows

v1.1-extended: the shipped artefacts and receipts only

v1.1: what ships is the original weights with the extended table and a raised prior; README says which file is which

v1.1: the vocabulary extended by 8,600 English pieces, the model taught to copy them; container, master weights, tables, receipts

pirate assistant: master weights, deployed container, the corpus and the generator

finetune kit: the master weights this run started from, alongside what it produced

finetune kit: the weights this run produced (date spans carry the year, 0/40 -> 40/40)

finetune kit: corpus builder, checker, trainer, the corpus used and the worked result

SHADOW 50M Instruct: one supervised pass from the 45B base; conversation 10/10, same-key 0.88, programs 10/40, golden 8/8

remove the stacked fine-tune weight and container, superseded by the single pass from the 45B base

model card: the .pt is the master weight of this run, with how to continue fine-tuning from it

SHADOW 50M Instruct: conversation, memory and record-selection fine-tune with replay; measured before/after

initial commit