Datasets:
Refresh Tabby-Pretrain TIME results: max(H, 128)
Updates results/Tabby-Pretrain using the released backbone at https://huggingface.co/paris-noah/Tabby-Pretrain (source revision 7fff2ae809bf49907f9dfaacce08092885fa6006). Zero-shot inference uses effective mask span max(H, 128), scores only the first H predictions, BF16, and an input-history cap of 4000 within the 8192-position model window. The backbone was trained for 165,000 steps. This PR contains 98 TIME task directories, each with config.json, metrics.npz, and predictions.npz. The prediction and metric files are byte-identical to the source run; this submission does not represent a new independent rerun. The companion Tabby prompt-model submission uses a different history cap (8096), so direct same-context comparisons would require a separate 8096 run for this backbone.