# Conversion changes

Original source: docling-project/ScreenVLM at 78df0cc8498bbea1a03eb2816056fc0a76d2f63d.
Converted original mixed F32/BF16 weights to GGUF F16 matrix storage with
required F32 tensors. Q8_0 is independently derived from F16. The full vision encoder and 4x connector are in mmproj.
No training was performed. GGUF metadata carries a prominent change notice.
The source tokenizer's unusable out-of-range IDs 100352 (<layout>) and
100353 (<end_of_utterance>) are removed only in a converter-specific copy.
Original source files and weights remain unchanged. No model-defined token
or weight row is removed; the source chat template is preserved verbatim.
This is an independent conversion and does not imply upstream endorsement.