warlockee/orze-nips-models

Orze NIPS 2026 — Collision Detection Models

Pre-trained checkpoints from the NeurIPS 2026 paper "Auto Research Is Not Auto Tuning: Convergence Analysis of 10,000 LLM-Guided Experiments".

These are frozen-feature temporal classifiers (Recipe A) trained on pre-extracted V-JEPA 2 / DINOv2 features from the Nexar Dashcam Collision Prediction dataset.

Models

Checkpoint	Test mAP	Params	Backbone
idea-502970	0.7853	~24MB	V-JEPA 2 frozen features
idea-eb79fc	0.7816	~11MB	V-JEPA 2 frozen features
idea-2c0263	0.7802	~24MB	V-JEPA 2 frozen features

Each directory contains:

• best_model.pt — PyTorch checkpoint (temporal classifier head)
• idea_config.yaml — Full training config (proposed by LLM agent)
• metrics.json — Training/val/test metrics

Note: These are the lightweight temporal heads that operate on pre-extracted backbone features. You must extract V-JEPA 2 / DINOv2 features from raw videos first using the scripts in the orze-nips repo.

Usage

from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(
    repo_id="warlockee/orze-nips-models",
    filename="idea-502970/best_model.pt"
)
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
# Load into temporal classifier defined in training/train.py

Reproduction

See the full codebase at https://github.com/warlockee/orze-nips

@inproceedings{anonymous2026autoresearch,
  title={Auto Research Is Not Auto Tuning: Convergence Analysis of 10,000 {LLM}-Guided Experiments},
  author={Anonymous},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}