nguyenviet21/AIC26-models

AIC runtime model bundle

This is a deployment bundle of existing model files, not a newly trained model. No competition video, image, transcript, embedding, submission, credentials or environment is included. bundle-manifest.json records file sizes and SHA-256. Use scripts/model_bundle.py restore from the companion GitHub code repository.

Models and provenance

Component	Source / identity
MetaCLIP2	https://huggingface.co/facebook/metaclip-2-worldwide-huge-quickgelu/tree/c139061af7b10fdb2e754b60d2b1182a3d5526c2
BGE-VL-large	https://huggingface.co/BAAI/BGE-VL-large/tree/40fb48217f521df22a2a5bf15edd52ed1146ef05
Perception Encoder	https://huggingface.co/facebook/PE-Core-B16-224/tree/a16450b46fef32363459920c2685a1b4ef13dcd9
BEiT-3 COCO retrieval base	https://github.com/microsoft/unilm/tree/833df7e7832e5064a281131ee64a481afa8e5b95/beit3 ; checkpoint identity is recorded by file hash
CRAFT	EasyOCR detector craft_mlt_25k.pth; https://github.com/JaidedAI/EasyOCR
PARSeq-VN	Team-provided best-parseq.ckpt; architecture https://github.com/baudm/parseq/tree/1572487d138cdf0340fb2a2f1437cc0cca05eabf
ChunkFormer	https://huggingface.co/khanhld/chunkformer-ctc-large-vie/tree/311fc03558a895dc2b32957f2fb4236c7fb1455b ; deployed local files are identified by their hashes
RF-DETR Medium	Local deployed rf-detr-medium.pth; https://github.com/roboflow/rf-detr ; checkpoint identity is recorded by file hash

Original authors retain their rights. MetaCLIP2 and ChunkFormer model cards state CC-BY-NC-4.0. Preserve upstream attribution and applicable notices; do not assign one blanket code license to this bundle. The publisher must add the provenance and redistribution terms for the team's PARSeq-VN checkpoint before publishing; the architecture's license alone does not establish the checkpoint's terms.

Layout

project/ contains files restored to their runtime paths inside the code repo. hub/ contains only selected pinned snapshots, restored to HF_HUB_CACHE under their original repository IDs. These are weights, model configs, tokenizer files and required BGE-VL Python loaders. Python environments and third-party source checkouts are installed separately from the GitHub instructions.

Optional Ollama qwen2.5:3b-instruct is provisioned separately with ollama pull. An Ollama model tag is not interchangeable with a Transformers checkpoint. External VLM is disabled in the baseline and has no selected weight in this bundle.