Reproducibility Artifacts for Synthetic Long-Term Preference Modeling in Multimodal Background Music Recommendation

This record contains reproducibility artifacts for the research project "A Unified Multimodal Large Language Model Based on Synthetic Long-Term Preference Modeling for Background Music Recommendation". The uploaded files include model checkpoints for the main experiments, final long-term preference cache files, preference vector files, metadata, split/index files, and feature-alignment lists used by the accompanying GitHub repository: [GitHub URL here]. File Contents `bgm_recommender_model_checkpoints_v1.tar.zst`: Best checkpoints for experiments `exp_01` through `exp_07`, including LoRA adapters, multimodal projectors, ranking heads, and related configuration files. `bgm_recommender_ltp_cache_v1.tar.zst`: Runtime cache files, final `preference_vectors*.h5` files, Stage 5 generation logs, and projection weights. `bgm_recommender_metadata_and_ids_v1.tar.zst`: Music metadata, HDF5 ID lists, split cache, and feature-alignment whitelist files. `CHECKSUMS.txt`: SHA256 checksums for verifying downloaded archives. Exclusions & Limitations Excluded Large / Raw Data: This record does not include the full MuseChat HDF5 feature folders (TB-scale), raw video/audio files, or the base LLaMA 2-7B model weights. Raw media files are excluded due to storage constraints and third-party licensing restrictions. Usage & Licensing: The base LLaMA model weights must be obtained separately from Meta/Hugging Face. Use of the released adapters, checkpoints, and derived artifacts is subject to the license terms of the base models (LLaMA 2, CLIP, AST) and the original MuseChat benchmark. Please refer to the accompanying GitHub repository for source code, setup documentation, experiment mapping, and instructions for restoring these files into the expected directory structure.