Refresh model card: add Pareto + architecture figures, TIME metrics, latency table

.gitattributes

@@ -34,3 +34,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 figures/architecture.png filter=lfs diff=lfs merge=lfs -text
+assets/architecture.png filter=lfs diff=lfs merge=lfs -text
+assets/pareto.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -15,58 +15,84 @@ thumbnail: https://corp.dd-static.net/img/about/presskit/kit/press_kit.png
 model-index:
 - name: Toto-2.0-22m
   results:
-    - task: 
+    - task:
         type: time-series-forecasting
-      dataset: 
-        name: GIFT-Eval
-        type: GIFT-Eval
+      dataset:
+        name: BOOM
+        type: BOOM
       metrics:
+        - name: CRPS
+          type: CRPS
+          value: 0.363
         - name: MASE
           type: MASE
-          value: 0.719
+          value: 0.601
+      source:
+        name: BOOM 💥 Observability Time-Series Forecasting Leaderboard
+        url: https://huggingface.co/spaces/Datadog/BOOM
+    - task:
+        type: time-series-forecasting
+      dataset:
+        name: GIFT-Eval
+        type: GIFT-Eval
+      metrics:
         - name: CRPS
           type: CRPS
           value: 0.496
+        - name: MASE
+          type: MASE
+          value: 0.719
       source:
         name: GIFT-Eval Time Series Forecasting Leaderboard
         url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
-  
-    - task: 
+    - task:
         type: time-series-forecasting
       dataset:
-        name: BOOM
-        type: BOOM
+        name: TIME
+        type: TIME
       metrics:
-        - name: MASE
-          type: MASE
-          value: 0.601
         - name: CRPS
           type: CRPS
-          value: 0.363
+          value: 0.556
+        - name: MASE
+          type: MASE
+          value: 0.668
       source:
-        name: BOOM 💥 Observability Time-Series Forecasting Leaderboard
-        url: https://huggingface.co/spaces/Datadog/BOOM
+        name: TIME Benchmark Leaderboard
+        url: https://huggingface.co/spaces/Real-TSF/TIME-leaderboard
 ---
+
 # Toto-2.0-22m
 
-Toto (**T**ime Series **O**ptimized **T**ransformer for [**O**bservability](https://www.datadoghq.com/knowledge-center/observability/)) is a family of time series foundation models for multivariate forecasting developed by [Datadog](https://www.datadoghq.com/). **Toto 2.0** is the current generation, featuring u-μP-scaled transformers ranging from 4M to 2.5B parameters.
+Toto (Time Series Optimized Transformer for [Observability](https://www.datadoghq.com/knowledge-center/observability/)) is a family of time series foundation models for multivariate forecasting developed by [Datadog](https://www.datadoghq.com/). Toto 2.0 is the current generation, featuring u-μP-scaled transformers ranging from 4M to 2.5B parameters, all trained from a single recipe. Forecast quality improves reliably with parameter count across the family, with no sign of saturation at 2.5B.
+
+The family sets a new state of the art on three forecasting benchmarks: [BOOM](https://huggingface.co/spaces/Datadog/BOOM), our observability benchmark; [GIFT-Eval](https://huggingface.co/spaces/Salesforce/GIFT-Eval), the standard general-purpose benchmark; and the recent contamination-resistant [TIME](https://arxiv.org/abs/2602.12147) benchmark.
 
 ---
 
 ## ✨ Key Features
 
-- **Zero-Shot Forecasting**: Forecast without fine-tuning on your specific time series.
-- **Multi-Variate Support**: Efficiently process multiple variables using alternating time/variate attention.
-- **Probabilistic Predictions**: Generate point forecasts and uncertainty estimates via a quantile output head.
-- **Decoder-Only Architecture**: Support for variable prediction horizons and context lengths.
-- **u-μP Scaling**: Stable training transfer across all model sizes.
+- **Zero-Shot Forecasting:** Forecast without fine-tuning on your specific time series.
+- **Multi-Variate Support:** Efficiently process multiple variables using alternating time/variate attention.
+- **Probabilistic Predictions:** Generate point forecasts and uncertainty estimates via a quantile output head.
+- **Decoder-Only Architecture:** Support for variable prediction horizons and context lengths.
+- **u-μP Scaling:** A single training recipe transfers cleanly across all five sizes (4M → 2.5B).
 
-<div style="width: 100%; margin: auto; padding: 1rem;">
-  <img src="figures/architecture.png" alt="Toto 2.0 architecture" style="width: 100%; height: auto;" />
-  <em style="display: block; margin-top: 0.5rem; text-align: center;">
-    Overview of the Toto 2.0 architecture.
-  </em>
-</div>
+---
+
+## 🏗️ Architecture
+
+![Overview of the Toto 2.0 architecture.](assets/architecture.png)
+
+A decoder-only patched transformer whose attention layers alternate between time-axis (causal) and variate-axis (full) views of the input. Toto 2.0 adds **contiguous patch masking (CPM)** for single-pass parallel decoding, a **quantile output head** trained with pinball loss, a robust arcsinh input scaler, residual MLP patch projections, and is trained with NorMuon. See the [technical report](#-additional-resources) for details.
+
+---
+
+## 📊 Performance
+
+![Pareto frontier on BOOM and GIFT-Eval](assets/pareto.png)
+
+Every Toto 2.0 size sits on or near the Pareto frontier on both BOOM and GIFT-Eval. The three largest sizes rank first, second, and third among foundation models on GIFT-Eval CRPS rank. On TIME, Toto 2.0 sizes take the top three spots on every metric, ahead of every other external foundation model evaluated.
 
 ---
 
@@ -87,18 +113,21 @@ import torch
 from toto2 import Toto2Model
 
 model = Toto2Model.from_pretrained("Datadog/Toto-2.0-22m")
-model = model.to("cuda").eval()
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = model.to(device).eval()
 
 # (batch, n_variates, time_steps)
-target = torch.randn(1, 1, 512, device="cuda")
+target = torch.randn(1, 1, 512, device=device)
 target_mask = torch.ones_like(target, dtype=torch.bool)
-series_ids = torch.zeros(1, 1, dtype=torch.long, device="cuda")
+series_ids = torch.zeros(1, 1, dtype=torch.long, device=device)
 
 # Returns quantiles of shape (9, batch, n_variates, horizon)
 # Quantile levels: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
 quantiles = model.forecast(
     {"target": target, "target_mask": target_mask, "series_ids": series_ids},
     horizon=96,
+    decode_block_size=768,
+    has_missing_values=False,
 )
 ```
 
@@ -108,22 +137,28 @@ For more examples, see the [Quick Start notebook](https://github.com/DataDog/tot
 
 ## 💾 Available Checkpoints
 
-| Checkpoint | Parameters |
-|---|---|
-| [Toto-2.0-4m](https://huggingface.co/Datadog/Toto-2.0-4m) | 4M |
-| [Toto-2.0-22m](https://huggingface.co/Datadog/Toto-2.0-22m) | 22M |
-| [Toto-2.0-313m](https://huggingface.co/Datadog/Toto-2.0-313m) | 313M |
-| [Toto-2.0-1B](https://huggingface.co/Datadog/Toto-2.0-1B) | 1B |
-| [Toto-2.0-2.5B](https://huggingface.co/Datadog/Toto-2.0-2.5B) | 2.5B |
+All five Toto 2.0 sizes share the same training recipe; pick a size based on your accuracy/latency budget. Latencies are forward-pass time for a 1,024-step forecast at batch size 8 on a single A100.
+
+| Model | Params | Single-pass latency<br>(1,024 horizon) | Block decoding<br>(block=768) | Recommended for |
+|---|---|---|---|---|
+| [Toto-2.0-4m](https://huggingface.co/Datadog/Toto-2.0-4m)     | 4m     | ~3.8 ms  | ~10.0 ms | Edge / CPU deployment; tightest latency or memory budgets. |
+| [Toto-2.0-22m](https://huggingface.co/Datadog/Toto-2.0-22m)   | 22m    | ~5.0 ms  | ~12.8 ms | Efficient default — matches or beats Toto 1.0 quality with ~7× fewer parameters. |
+| [Toto-2.0-313m](https://huggingface.co/Datadog/Toto-2.0-313m) | 313m   | ~15.4 ms | ~32.4 ms | Strong general-purpose checkpoint; top-3 foundation model on GIFT-Eval. |
+| [Toto-2.0-1B](https://huggingface.co/Datadog/Toto-2.0-1B)     | 1B     | ~20.9 ms | ~46.3 ms | Best quality / cost tradeoff for production workloads. |
+| [Toto-2.0-2.5B](https://huggingface.co/Datadog/Toto-2.0-2.5B) | 2.5B   | ~36.2 ms | ~78.0 ms | Highest accuracy; #1 foundation model on every benchmark. |
+
+> Single-pass decoding fills the entire horizon in one forward pass and is recommended up to ~768 steps. Block decoding generates the horizon in 768-step segments conditioned on the previous segment's median (with KV caching); it is slower but more stable at long horizons. Both modes use the same checkpoint.
 
 ---
 
 ## 🔗 Additional Resources
 
-- **[Blog Post](https://www.datadoghq.com/blog/ai/toto-2/)**
-- **[GitHub Repository](https://github.com/DataDog/toto)**
-- **[BOOM Dataset](https://huggingface.co/datasets/Datadog/BOOM)**
-- **[Toto 1.0 Weights](https://huggingface.co/Datadog/Toto-Open-Base-1.0)**
+- **Technical Report** — *(coming soon)*
+- [Blog Post](https://www.datadoghq.com/blog/ai/toto-2/)
+- [GitHub Repository](https://github.com/DataDog/toto)
+- [Toto 2.0 Collection](https://huggingface.co/collections/Datadog/toto-20) — all five base checkpoints
+- [BOOM Dataset](https://huggingface.co/datasets/Datadog/BOOM) — Datadog's observability time-series benchmark
+- [Toto 1.0 Weights](https://huggingface.co/Datadog/Toto-Open-Base-1.0)
 
 ---