Re-export with consistent stop tokens (0_5b first)
#1
by msluszniak - opened
The two published variants of each size disagree on where generation stops: bf16 carries get_eos_ids [151643] (<|endoftext|>) and 8da4w carries [151645] (<|im_end|>). The config matched only the bf16 half, and that half is the broken one -- those builds emit <|im_end|> as visible text and keep generating.
All files are re-exported to carry upstream's [151645, 151643]. Configs additionally gain get_n_layers (null in all three) and the real 2047 sequence dim.
0_5b lands in this commit; 1_5b and 3b follow as further commits to this PR, since all six files do not fit on disk at once. Each .pte is probed against its config before upload.