Gramscii-IT/SemanticRepair-270M

Card: the one wrong execution, named — recovered by re-running the seeded bench

Card: the licence duties by section, the compound bound, the surface's last byte

The linter that reports collisions between descriptions, which is the other half of what you can improve

What bounds the weaknesses: nothing executed, the source message kept beside, the bars, and the descriptions

Link the base model where it is named

Every column the bench reports, including the two the training did not buy

Published by the organisation, not a personal account

Tables where there were paragraphs: the machine, the run, the generators, the two roles

The full machine, because M4 Pro alone does not say which one

One desktop machine, and nothing left it: the hardware, the ninety minutes, the 7.3 GB

Name the models that wrote the data and the judge that refused 72,996 rows, including where judge and generator stopped being different

What it does on CPU alone, measured, and the three things that measurement does not settle about a phone

Tags people actually search: query-understanding, intent-detection, rag, slm

The exact call the engine makes, and the date figure re-measured on these weights instead of quoted from another checkpoint

What the two injection families actually teach, counted rather than inferred from their names

Name the data: where it is, how big, how it splits by language and by what each message is doing

Measure against the base and against no repair, not against versions nobody can download

The fused tokenizer declares its own special tokens; this one came from the base model

The fused tokenizer, the real outputs, and the limits the first card did not name

Remove the GRPO checkpoint: it is not what is served, and it never passed its own pre-registered rule

The model the engine actually serves: supervised v19, with the benches behind it

The card says the data is not published yet, and why

The Q8_0 the seat serves, byte-identical to the pinned file

model.safetensors.index.json

model.safetensors

chat_template.jinja

tokenizer.json

tokenizer_config.json

generation_config.json

config.json

The card states the surface, the training and what it gets wrong

initial commit