Tabby — seed 42, 128-step minimum forecast mask

This is a self-contained Tabby checkpoint: the frozen Tabby-Pretrain backbone plus a seed-42 prompt. Its 99 quantile forecasts cover levels 0.01 through 0.99. The backbone weights come from pretraining step 165,000. The prompt was trained with the backbone frozen and selected at epoch 14 (val_loss=2.0483945271).

Resources

• Open in Colab
• TabbyTSFM GitHub code
• Technical report

Inference rule

For a requested horizon H, the model masks max(H, 128) future positions and returns the first H forecasts from that span. The backbone window is 8,192 positions. Consequently, the most observed history it can use is 8192 - max(H, 128), even when the evaluation input cap is 8,096. For example, H=96 uses at most 8,064 history points; H=168 uses at most 8,024.

The original seed-42 training checkpoint records min_forecast_span=0. This release changes only the inference setting to 128. All 25 prompt tensors and all 253 backbone tensors are unchanged. config.json records min_forecast_span=128, so ordinary predict calls use the released rule without a separate flag.

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("paris-noah/Tabby", trust_remote_code=True).eval()
history = torch.randn(2, 8096).cumsum(-1)
quantiles = model.predict(history, prediction_length=96)  # [2, 99, 96]
median = quantiles[:, 49]  # 0.50 quantile

The model forecasts each variate separately. It accepts missing history values as NaN and performs its own visible-history normalization.

Architecture and training

The frozen backbone has 20 layers, width 768, 12 attention heads, and 16-step patches. The full release has 146.7 million parameters, including a 160-token prompt module. The prompt uses 16 adaptive context segments, a rank-4 context map, and a cross-attention refiner. Only the prompt was adapted after pretraining, using 48 GIFT-Eval training tasks. The released prompt has seed 42 and was selected at epoch 14; the backbone remains frozen.

GIFT-Eval results

Seasonal-Naive-normalized geometric means across the same 97 configurations, with an 8,096-point input history cap and the minimum mask span of 128:

These results come from the supplied GIFT-Eval evaluation CSVs. Prompt post-training used GIFT-Eval training tasks; the corresponding leaderboard submission marks testdata_leakage as Yes. The prompt checkpoint is a new seed and the inference rule also differs from the previous seed-4 Tabby release, so their score difference does not isolate either change.

The accompanying TIME result archive contains 98 task outputs at an 8,096-point input cap for this prompted model. Its public leaderboard aggregate should be computed by TIME's official pipeline when the results are submitted.

Files and provenance

• model.safetensors: 278 tensors (253 backbone, 25 prompt), 146.7 million parameters.
• config.json: architecture, prompt settings, backbone revision, and 128-step minimum mask span.
• modeling_tabby.py, configuration_tabby.py, and their local modules: the existing Hugging Face Tabby inference implementation, which already supports this mask rule.

Code: TabbyTSFM. License: CC BY-NC 4.0.