Oratis's picture
创建实验模型公开归档并说明来源、许可与导出约束
2feee1b verified
Raw History Blame Contribute Delete
1.54 kB
Diogenes / CuddlerAI — historical research-model archive, 2026-11-15.
Own contributions to these archived research-model artifacts: CC BY-NC 4.0.
https://creativecommons.org/licenses/by-nc/4.0/
Base models (full weights not included):
- Qwen Team, Qwen/Qwen3-VL-4B-Instruct, Apache-2.0.
https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct
- Meta AI, DINOv2 ViT-S/14, Apache-2.0.
https://github.com/facebookresearch/dinov2
https://github.com/facebookresearch/dinov2/blob/main/LICENSE
Data lineage (downloaded source images/videos not included):
- Markov AI, gaming-500-hours, current card CC BY-NC 4.0.
https://huggingface.co/datasets/markov-ai/gaming-500-hours
- Open World Agents, D2E-480p, current card CC BY-NC 4.0.
https://huggingface.co/datasets/open-world-agents/D2E-480p
D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI.
https://arxiv.org/abs/2510.05684
Changes include method-validation fine-tuning, LoRA deltas, trained IDM heads,
benchmark conversion and teacher annotation in the historical experiments.
Closeout extracts original LoRA files and exports trained head tensors to
Safetensors without altering tensor bytes; it removes the frozen downloaded
trunk from full checkpoints. Per-variant manifests record the changes and hashes.
Historical base and data revisions are unknown; current references are not proof
of the exact historical inputs. Original method-validation/non-trainable status
remains unchanged. Third-party game rights are not granted by this archive.