Tabby-Prompt — seed 42, inference mask 128

This optional prompt-only package contains the 25 prompt tensors for the frozen Tabby-Pretrain backbone at pretraining step 165,000. The prompt has 160 tokens, uses segmented context-aware conditioning, and was selected at epoch 14 (val_loss=2.0483945271).

The earlier tabby_prompt_plen160_seed4.pt remains in this repository for historical reproducibility. Use tabby_prompt_plen160_seed42_mask128.pt for the current release.

tabby_prompt_plen160_seed42_mask128.pt has identical tensor values to the supplied seed-42 checkpoint_best.pt. Its inference config sets min_forecast_span=128, making the existing TabbyTSFM GIFT-Eval and TIME prompt evaluators mask max(H, 128) positions and score the first H. The original training setting was min_forecast_span=0, retained as training_min_forecast_span=0 for provenance. The internal pretraining path was replaced with the public backbone reference.

The full, independently loadable backbone-plus-prompt release is the Tabby model package. This prompt-only file is for reproducing evaluations through TabbyTSFM or for prompt research.

For GIFT-Eval, use --mode prompt --ckpt tabby_prompt_plen160_seed42_mask128.pt --pretrain_ckpt <Tabby-Pretrain snapshot> --context_length 8096 --precision bf16 with benchmarks/forecasting/gift_eval/evaluate.py. For TIME, use the same checkpoint and backbone with benchmarks/forecasting/time/evaluate.py --mode prompt --context_length 8096 and the official TIME data. The checkpoint's 128 setting is read automatically in prompt mode.

The GIFT-Eval 97-configuration Seasonal-Naive-normalized geometric means are MASE 0.6880 and CRPS 0.4758 under this inference protocol. Prompt post-training used GIFT-Eval training tasks.

License: CC BY-NC 4.0.