Spaces:
Running
Running
| title: Saccadic | IRIS Training Observations | |
| emoji: 👁 | |
| colorFrom: purple | |
| colorTo: blue | |
| sdk: static | |
| app_file: index.html | |
| pinned: false | |
| short_description: A closer look inside neural network training with IRIS | |
| <img src="https://cdn-avatars.huggingface.co/v1/production/uploads/69d6d16790de87a3b534bf99/SkapdHGpRMem-dCvIvgMx.png" alt="Saccadic purple lotus with a blue eye" width="110"> | |
| # IRIS™ · A closer look at AI training | |
| *A better view while the run is still underway.* | |
| Saccadic LLC develops instruments for observing neural network training. IRIS, the Intrinsic Rank Inspection System, adds a measured view of internal structure, specialization, and coordination alongside loss and validation curves. | |
| [Explore the styled IRIS Space](https://huggingface.co/spaces/SaccadicAI/README) · [IRIS product details](https://www.saccadic.ai/iris) · [Discuss an evaluation](mailto:info@saccadic.ai) | |
| ## Look inside the run | |
| - **Flow Map:** inspect measured connections between layers and experts for different input domains. | |
| - **Representation structure:** observe how activation variation concentrates and how dominant directions move. | |
| - **Expert routing:** compare which subjects reach which experts, separately for each layer. | |
| - **Domain interaction:** inspect sharing and training-gradient agreement alongside loss and task results. | |
| ## Capture. Record. Inspect. | |
| The tracker captures measurements using consistent probe inputs. A self-contained **.iris** file records the run. The local **Bench** reads that growing record during training and opens it afterward without loading the model. | |
| Each graph has an Explain guide. Teams can define labeled scalar signals, derive formulas from captured channels, and export custom charts. | |
| ## Follow the signal through the model | |
|  | |
| *Figure 10.1. A snapshot during training at step 49,300. Green links show mechanics; orange links show arithmetic. Line thickness represents measured signal, and expert circles are sized by traffic share. This is a static image of the IRIS Flow Map.* | |
| The map provides an internal view alongside the accuracy curves: which experts carry each domain’s measured signal, where paths overlap, and how traffic is distributed across layers. Revisit these measurements across the recorded run to inspect how the patterns change. | |
| ## A measured example | |
|  | |
| *Figure 4.4. Training and held-out accuracy for the mathematical operations in this measured run. [Read the experiment and its conditions.](https://huggingface.co/blog/Saccadic-AI/accuracy-plateaued-the-model-kept-changing)* | |
| ## Work in your own environment | |
| IRIS fits existing PyTorch training loops, with capture and inspection on CPU, Apple MPS, or NVIDIA CUDA. Coverage depends on model layout; shared-weight configurations are currently unsupported. | |
| Model weights, activations, and probe text are not sent to Saccadic. Licensing and metering contact Saccadic. Optional configured relays send selected scalar measurements to existing experiment trackers. | |
| ## One model. One run. One concrete investigation. | |
| We welcome research and joint evaluations. Academic researchers can request credit grants; requests are reviewed individually. See the public product site for availability and licensing terms. | |
| [info@saccadic.ai](mailto:info@saccadic.ai) · [www.saccadic.ai/iris](https://www.saccadic.ai/iris) | |
| *IRIS pre-alpha. Product overview based on the presentation dated 24 October 2026. IRIS records observations; interpret readings alongside gradients, loss, validation, and task results.* | |