---
language:
- en
tags:
- recommendation
- multimodal
- sequential-recommendation
---

# MoSID preprocessed data

Preprocessed item features and sequence splits for the Movies dataset in nm mode.

## Layout

The files under `Movies/nm/` use the same contiguous item IDs. ID 0 is reserved for padding; valid item IDs are 1 through N.

| File | Content |
| --- | --- |
| txt_emb.pt | Text embedding tensor, shape [N+1, 768] |
| img_emb.pt | Visual embedding tensor, shape [N+1, 768] |
| collab_emb.pt | Collaborative embedding tensor, shape [N+1, 256] |
| mod_presence.pt | Boolean tensor, shape [N+1, 2], for text and visual availability |
| mappings.pkl | Mapping dictionary containing item2id |
| train.pkl | Training item_seqs and labels |
| valid.pkl | Validation item_seqs and labels |
| test.pkl | Test item_seqs and labels |

Place the `Movies` directory under `datasets/amazon/` in the MoSID source repository. Semantic IDs and codebook weights are generated by the Stage 1 training and export commands provided in the source repository.

## Integrity

`SHA256SUMS.txt` contains the SHA-256 hashes of the released data files.

## Serialization

The `.pt` files contain PyTorch tensors. The `.pkl` files use Python pickle. Only load serialized files from a trusted source.