File size: 2,235 Bytes
989b97a
 
 
 
 
2df5234
 
 
 
989b97a
 
 
 
2df5234
989b97a
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
---
library_name: pytorch
license: other
license_name: per-asset-terms
tags:
- scientific-machine-learning
- model-editing
- safetensors
- depth-estimation
---

# HypoLoop

Native model weights and registered feedback data from [**HypoLoop: Revising Weight-Editing Plans from Scientific Feedback**](https://huggingface.co/papers/2611.23718). The implementation, preprint, and loading instructions are available on GitHub. [HypoLoop repository](https://github.com/limengran98/HypoLoop) 路 [Loading guide](https://github.com/limengran98/HypoLoop/blob/main/docs/HUGGING_FACE.md)

| Paper task | Data | Model weights |
| :--- | :--- | :--- |
| PHY01 路 Burgers | Registered field histories, spatial grid, feedback and held-out targets | FNO1d base and final fixed- and dynamic-location checkpoints for both services |
| MATH01 路 Constrained QP | Registered right-hand sides, objective and constraint matrices, feedback and held-out solutions | DC3 base and final fixed- and dynamic-location checkpoints for both services |
| DEP02 路 DIODE | 24 registered RGB images, scene IDs, sampled depth targets and grid coordinates | Fixed-plan and revised-plan Claude head replacements for Depth Anything V2 Metric VKITTI ViT-S |
| DEP01 路 Make3D | [Original Make3D dataset](https://cs.stanford.edu/people/asaxena/make3d/) | DeepSeek HL-Global head replacements for the same Depth Anything V2 base |

[manifest.json](manifest.json) lists every checkpoint, its architecture, and the corresponding data files. [checksums.sha256](checksums.sha256) records the upload contents. [LICENSES.md](LICENSES.md) lists source attribution and terms by asset.

The registered order is initial feedback `F0`, subsequent feedback `B1` and `B2`, then held-out evaluation `E`. Use `E` for evaluation. Each target archive aligns `ids` with `y`. DIODE uses the 41脳61 sampled grid in [grid.json](data/diode/grid.json), with `NaN` for invalid depth. The 12 reference images are from the original Make3D training collection.

The two depth cases use named-tensor replacements. Load the original VKITTI ViT-S metric checkpoint before applying a replacement. [Depth Anything V2 upstream model](https://huggingface.co/depth-anything/Depth-Anything-V2-Metric-VKITTI-Small)